{
 "meta": {
  "title": "Strim",
  "machine": "Apple M5 Max, 64 GB, macOS 27.0.1",
  "model": "Qwen3.8-27B 4-bit (models/qwen38-4bit) + MTP head (models/qwen38-mtp-4bit)",
  "generated": "final.* figures are generated from bench/results/h2h-final3-20261028 by strim site build (re-parsed and cross-checked on every run); the others are hand-entered and strim site check verifies each value (verify string, verify_count or table cell) against its source file."
 },
 "datasets": [
  {
   "id": "final",
   "label": "Final head-to-head, Strim vs rapid-mlx vs mlx-vlm",
   "path": "bench/results/h2h-final3-20261028",
   "needs": "summary.md"
  },
  {
   "id": "soak",
   "label": "30-minute real-client soak",
   "path": "bench/results/soak-20261027",
   "kind": "soak"
  },
  {
   "id": "harness",
   "label": "Agent harness and MCP status",
   "path": "bench/results/mcp-clients-20261027",
   "kind": "hand-entered"
  }
 ],
 "figures": [
  {
   "id": "pc.4k.miss",
   "value": 3.994,
   "unit": "s",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-4096.log",
   "method_note": "checkprefix N=4096, shared 3072 tokens: TTFT on a miss (no store).",
   "caveat": "Synthetic prompts, in-process, no HTTP. Idle GPU assumed; load not recorded.",
   "n": "1"
  },
  {
   "id": "pc.4k.hit",
   "value": 1.022,
   "unit": "s",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-4096.log",
   "method_note": "checkprefix N=4096, shared 3072 tokens: TTFT with a 3072-token prefix hit.",
   "caveat": "Synthetic prompts, in-process, no HTTP.",
   "n": "1"
  },
  {
   "id": "pc.4k.speedup",
   "value": 3.91,
   "unit": "x",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-4096.log",
   "method_note": "miss / hit TTFT as printed by checkprefix.",
   "caveat": "Speedup grows with the shared fraction; here 3072 of 4096 tokens shared.",
   "verify": "3.91x"
  },
  {
   "id": "pc.8k_6k.miss",
   "value": 8.204,
   "unit": "s",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-8192.log",
   "method_note": "checkprefix N=8192, shared 6144: TTFT on a miss.",
   "caveat": "Synthetic prompts, in-process.",
   "n": "1"
  },
  {
   "id": "pc.8k_6k.hit",
   "value": 3.44,
   "unit": "s",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-8192.log",
   "method_note": "checkprefix N=8192, shared 6144: TTFT on a hit.",
   "caveat": "Synthetic prompts, in-process. Recomputes the 2048-token tail.",
   "n": "1",
   "verify": "3.440",
   "dp": 3
  },
  {
   "id": "pc.8k_6k.speedup",
   "value": 2.38,
   "unit": "x",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-8192.log",
   "method_note": "miss / hit TTFT as printed by checkprefix.",
   "caveat": "",
   "verify": "2.38x"
  },
  {
   "id": "pc.mtp4k.miss",
   "value": 4.094,
   "unit": "s",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "checkprefix with MTP head, N=4096, shared 3072: TTFT (prefill + seed) on a miss.",
   "caveat": "In-process; MTP head state is stored in the trie nodes.",
   "n": "1"
  },
  {
   "id": "pc.mtp4k.hit",
   "value": 1.051,
   "unit": "s",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "checkprefix with MTP head, N=4096, shared 3072: TTFT (prefill + seed) on a hit.",
   "caveat": "In-process.",
   "n": "1"
  },
  {
   "id": "pc.mtp4k.speedup",
   "value": 3.9,
   "unit": "x",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "miss / hit TTFT with MTP over the cache.",
   "caveat": "",
   "verify": "3.90x",
   "dp": 2
  },
  {
   "id": "pc.mtp8k.miss",
   "value": 8.157,
   "unit": "s",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-8192-mtp.log",
   "method_note": "checkprefix N=8192, shared 3072: TTFT on a miss.",
   "caveat": "Synthetic prompts, in-process.",
   "n": "1"
  },
  {
   "id": "pc.mtp8k.hit",
   "value": 1.091,
   "unit": "s",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-8192-mtp.log",
   "method_note": "checkprefix N=8192, shared 3072: TTFT on a hit.",
   "caveat": "Synthetic prompts, in-process. The log header says shared=3072, but the store at that step already held the full prompt's nodes (21 nodes, 4427 MB), so the matched length is probably larger than 3072 and is not printed for this step. The log's section (g) 8K numbers (hit 6.084 s vs miss 22.906 s) were taken at a much slower prefill rate (367 tok/s, versus about 1,000 elsewhere) and are not used.",
   "n": "1",
   "verify": "1.091"
  },
  {
   "id": "pc.mtp8k.speedup",
   "value": 7.48,
   "unit": "x",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-8192-mtp.log",
   "method_note": "miss / hit TTFT as printed by checkprefix.",
   "caveat": "See pc.mtp8k.hit caveat.",
   "verify": "7.48x"
  },
  {
   "id": "pc.node_mb",
   "value": 210.8,
   "unit": "MB",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "Plain trie node (1024 tokens) accounted size.",
   "caveat": "Accounted size; see memory section for the real-memory finding that preceded the conv-state fix.",
   "verify": "210.8"
  },
  {
   "id": "pc.mtp_node_extra_mb",
   "value": 4.01,
   "unit": "MB",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "Extra bytes per node for the MTP head state (+1.90%).",
   "caveat": "",
   "verify": "4.01"
  },
  {
   "id": "pc.policy_cost_4k",
   "value": 0.6,
   "unit": "%",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-4096.log",
   "method_note": "Cost of the aligned 1024-token chunk prefill policy versus one-shot prefill at N=4096 (chunked 1026 tok/s, one-shot 1032 tok/s).",
   "caveat": "Single measurement.",
   "verify": "0.6%"
  },
  {
   "id": "pc.mtp_accept_4k",
   "value": 65,
   "unit": "%",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "MTP draft acceptance in the 4K cache-hit spec-decode check (2.37 tokens/pass; hit 48.3 vs miss 49.9 tok/s).",
   "caveat": "64 tokens on one synthetic prompt; not representative of real text.",
   "verify": "65%"
  },
  {
   "id": "pc.hot_continue",
   "value": "IDENTICAL over 20 tokens",
   "unit": "",
   "source_path": "bench/results/dec-wave3-20261026/checkprefix-4096-mtp.log",
   "method_note": "Hot session continuation vs cold re-prefill, 20 greedy tokens.",
   "caveat": "Informational check; min cold margin 0.000 (a bf16 tie was present and still matched).",
   "verify": "IDENTICAL over 20 tokens"
  },
  {
   "id": "lc.p4k_a.serial",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p4k_a.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "16/16"
  },
  {
   "id": "lc.p4k_a.tensor",
   "value": "14/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p4k_a.",
   "caveat": "Same near-tie caveat.",
   "verify": "14/16"
  },
  {
   "id": "lc.p4k_b.serial",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p4k_b.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "16/16"
  },
  {
   "id": "lc.p4k_b.tensor",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p4k_b.",
   "caveat": "Same near-tie caveat.",
   "verify": "16/16"
  },
  {
   "id": "lc.p4k_c.serial",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p4k_c.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "16/16"
  },
  {
   "id": "lc.p4k_c.tensor",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p4k_c.",
   "caveat": "Same near-tie caveat.",
   "verify": "16/16"
  },
  {
   "id": "lc.p4k_d.serial",
   "value": "6/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p4k_d.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "6/16"
  },
  {
   "id": "lc.p4k_d.tensor",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p4k_d.",
   "caveat": "Same near-tie caveat.",
   "verify": "16/16"
  },
  {
   "id": "lc.p4k_e.serial",
   "value": "4/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p4k_e.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "4/16"
  },
  {
   "id": "lc.p4k_e.tensor",
   "value": "4/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p4k_e.",
   "caveat": "Same near-tie caveat.",
   "verify": "4/16"
  },
  {
   "id": "lc.p1k_a.serial",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p1k_a.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "16/16"
  },
  {
   "id": "lc.p1k_a.tensor",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p1k_a.",
   "caveat": "Same near-tie caveat.",
   "verify": "16/16"
  },
  {
   "id": "lc.p1k_b.serial",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p1k_b.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "16/16"
  },
  {
   "id": "lc.p1k_b.tensor",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p1k_b.",
   "caveat": "Same near-tie caveat.",
   "verify": "16/16"
  },
  {
   "id": "lc.p1k_c.serial",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs serial GDN kernel, greedy 16 tokens, prompt p1k_c.",
   "caveat": "Divergence is at a bf16 near-tie (logits printed bf16-rounded); a single tie flips all later tokens.",
   "verify": "16/16"
  },
  {
   "id": "lc.p1k_c.tensor",
   "value": "16/16",
   "unit": "tokens agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Blocked (default) vs tensor-recurrence prefill, greedy 16 tokens, prompt p1k_c.",
   "caveat": "Same near-tie caveat.",
   "verify": "16/16"
  },
  {
   "id": "lc.first_token",
   "value": "8/8",
   "unit": "prompts agree",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "First generated token agrees across default, serial and tensor paths.",
   "caveat": "Natural repo prose, disjoint windows, Qwen tokenizer.",
   "verify": "8/8"
  },
  {
   "id": "lc.spec_exact_4k",
   "value": "0/2",
   "unit": "EXACT",
   "source_path": "bench/results/longctx-validate-20261026/summary.md",
   "method_note": "Speculative vs greedy parity at 4K: p4k_a and p4k_c both MISMATCH (first diff at generated token 14 and 18).",
   "caveat": "Not EXACT at 4K; same as the HANDOFF note.",
   "verify": "2/2"
  },
  {
   "id": "bm.ms_b1_s1",
   "value": 34.0,
   "unit": "ms/pass",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Verify-pass cost, B=1, S=1, 1K context.",
   "caveat": "Median of 6 rounds; background load avg 3-4, absolute ms +-5%.",
   "verify": "34.0"
  },
  {
   "id": "bm.ms_b1_s2",
   "value": 35.1,
   "unit": "ms/pass",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Verify-pass cost, B=1, S=2, 1K context.",
   "caveat": "As above.",
   "verify": "35.1"
  },
  {
   "id": "bm.ms_b4_s1",
   "value": 56.0,
   "unit": "ms/pass",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Verify-pass cost, B=4, S=1, 1K context.",
   "caveat": "As above.",
   "verify": "56.0"
  },
  {
   "id": "bm.ms_b4_s3",
   "value": 120.3,
   "unit": "ms/pass",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Verify-pass cost, B=4, S=3, 1K context.",
   "caveat": "As above.",
   "verify": "120.3"
  },
  {
   "id": "bm.bound.b1.prose_s2",
   "value": 1.46,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=1, prose s2.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.46x"
  },
  {
   "id": "bm.bound.b1.prose_s3",
   "value": 1.79,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=1, prose s3.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.79x"
  },
  {
   "id": "bm.bound.b1.code_s4",
   "value": 3.13,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=1, code s4.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "3.13x"
  },
  {
   "id": "bm.bound.b2.prose_s2",
   "value": 1.25,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=2, prose s2.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.25x"
  },
  {
   "id": "bm.bound.b2.prose_s3",
   "value": 1.2,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=2, prose s3.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.20x"
  },
  {
   "id": "bm.bound.b2.code_s4",
   "value": 1.9,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=2, code s4.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.90x"
  },
  {
   "id": "bm.bound.b3.prose_s2",
   "value": 1.03,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=3, prose s2.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.03x"
  },
  {
   "id": "bm.bound.b3.prose_s3",
   "value": 1.06,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=3, prose s3.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.06x"
  },
  {
   "id": "bm.bound.b3.code_s4",
   "value": 1.55,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=3, code s4.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.55x"
  },
  {
   "id": "bm.bound.b4.prose_s2",
   "value": 0.99,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=4, prose s2.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "0.99x"
  },
  {
   "id": "bm.bound.b4.prose_s3",
   "value": 0.93,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=4, prose s3.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "0.93x"
  },
  {
   "id": "bm.bound.b4.code_s4",
   "value": 1.57,
   "unit": "x",
   "source_path": "docs/BATCH-MTP-2026-10-26.md",
   "method_note": "Upper-bound aggregate speedup of in-batch MTP over plain batched decode, B=4, code s4.",
   "caveat": "Optimistic: ignores draft cost (2-6 ms/pass), reseed and rollback. Verdict: NO-GO for general BatchSpec.",
   "verify": "1.57x"
  },
  {
   "id": "ps.qmm_ratio",
   "value": "0.90-0.98",
   "unit": "qmm / bf16",
   "source_path": "docs/PREFILL-SPIKE-2026-10-26.md",
   "method_note": "MLX 4-bit qmm throughput as a fraction of dense bf16 matmul, 12 shape/M combinations.",
   "caveat": "MLX already dispatches NAX kernels on this machine.",
   "verify": "0.90"
  },
  {
   "id": "ps.plateau",
   "value": "62-63",
   "unit": "TF/s",
   "source_path": "docs/PREFILL-SPIKE-2026-10-26.md",
   "method_note": "Sustained dense GEMM plateau across shapes and dtypes.",
   "caveat": "True hardware peak unverified.",
   "verify": "62-63"
  },
  {
   "id": "ps.bound",
   "value": "+5-8%",
   "unit": "end-to-end",
   "source_path": "docs/PREFILL-SPIKE-2026-10-26.md",
   "method_note": "Upper bound on end-to-end 4K prefill gain from a custom 4-bit GEMM (about 990-1,020 tok/s from 944).",
   "caveat": "Optimistic bound; gate was 1.3x, step 3 not built.",
   "verify": "+5-8%"
  },
  {
   "id": "ps.prefill_4k",
   "value": 944,
   "unit": "tok/s",
   "source_path": "docs/PREFILL-SPIKE-2026-10-26.md",
   "method_note": "benchprefill 4096, single pass, rep 3 (936 / 947 / 944).",
   "caveat": "Loaded machine (load 2.7-5.0).",
   "verify": "944"
  },
  {
   "id": "ps.gemm_share",
   "value": 90,
   "unit": "%",
   "source_path": "docs/PREFILL-SPIKE-2026-10-26.md",
   "method_note": "Share of 4K prefill that is GEMM-bound (about 90-94%).",
   "caveat": "Derived from effective TF/s, not a profile.",
   "verify": "~90%"
  },
  {
   "id": "bg.before",
   "value": "825-846",
   "unit": "tok/s",
   "source_path": "docs/BLOCKED-GDN-PREFILL-2026-10-26.md",
   "method_note": "4K prefill with the serial GDN kernel.",
   "caveat": "Light and heavy load mixed.",
   "verify": "825-846"
  },
  {
   "id": "bg.after_light",
   "value": 915,
   "unit": "tok/s",
   "source_path": "docs/BLOCKED-GDN-PREFILL-2026-10-26.md",
   "method_note": "4K prefill with the blocked GDN kernel, load 3.7.",
   "caveat": "Kernel ported from rapid-mlx (Apache-2.0).",
   "verify": "915"
  },
  {
   "id": "bg.after_heavy",
   "value": 878,
   "unit": "tok/s",
   "source_path": "docs/BLOCKED-GDN-PREFILL-2026-10-26.md",
   "method_note": "4K prefill with the blocked GDN kernel, load 7.3.",
   "caveat": "Kernel ported from rapid-mlx (Apache-2.0).",
   "verify": "878"
  },
  {
   "id": "bp.b4_4k",
   "value": 0.95,
   "unit": "x",
   "source_path": "docs/BATCH-PREFILL-2026-10-26.md",
   "method_note": "Batched vs serial 4K prefill, B=4: slower.",
   "caveat": "Not adopted.",
   "verify": "0.95x"
  },
  {
   "id": "soak.smoke.tasks",
   "value": "2/2",
   "unit": "tasks passed",
   "source_path": "bench/results/soak-20261027/smoke-20261026T212306Z/summary.txt",
   "method_note": "5-minute smoke run (one claude -p task, one codex exec task).",
   "caveat": "n=2 only; shows the harness path works, not reliability.",
   "verify": "2 total, 2 passed"
  },
  {
   "id": "soak.smoke.crashes",
   "value": 0,
   "unit": "",
   "source_path": "bench/results/soak-20261027/smoke-20261026T212306Z/summary.txt",
   "method_note": "Server crashes in the 5-minute smoke run.",
   "caveat": "n=2 tasks, 0.6 min.",
   "verify": "crashes: 0"
  },
  {
   "id": "soak.tasks_total",
   "value": 120,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Tasks run. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. Pass = deterministic check + client exit 0 within the cap; failures are listed in the summary file.",
   "dataset": "soak",
   "verify": "120"
  },
  {
   "id": "soak.tasks_passed",
   "value": 116,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Tasks passed (check + client exit 0). 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. Pass = deterministic check + client exit 0 within the cap; failures are listed in the summary file.",
   "dataset": "soak",
   "verify": "116"
  },
  {
   "id": "soak.pass_claude",
   "value": 93.3,
   "unit": "%",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Claude Code pass rate. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. Pass = deterministic check + client exit 0 within the cap; failures are listed in the summary file.",
   "dataset": "soak",
   "verify": "93.3"
  },
  {
   "id": "soak.pass_codex",
   "value": 100.0,
   "unit": "%",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Codex pass rate. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. Pass = deterministic check + client exit 0 within the cap; failures are listed in the summary file.",
   "dataset": "soak",
   "verify": "100.0"
  },
  {
   "id": "soak.crashes",
   "value": 0,
   "unit": "",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Server crashes. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "0"
  },
  {
   "id": "soak.restarts",
   "value": 0,
   "unit": "",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Server restarts. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "0"
  },
  {
   "id": "soak.http5xx",
   "value": 0,
   "unit": "",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "HTTP 5xx events. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "0"
  },
  {
   "id": "soak.requests_total",
   "value": 665,
   "unit": "",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Server requests served. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "665"
  },
  {
   "id": "soak.requests_failed",
   "value": 0,
   "unit": "",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Server requests failed. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "0"
  },
  {
   "id": "soak.rss_slope",
   "value": 24.0,
   "unit": "MB/h",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "RSS slope over the run. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "24.0"
  },
  {
   "id": "soak.ttft_q1",
   "value": 249,
   "unit": "ms",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "TTFT median, first quartile of requests. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "249"
  },
  {
   "id": "soak.ttft_q4",
   "value": 310.5,
   "unit": "ms",
   "dp": 0,
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "TTFT median, last quartile of requests. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "310.5"
  },
  {
   "id": "soak.ttft_p90",
   "value": 1334,
   "unit": "ms",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "TTFT p90. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "1334"
  },
  {
   "id": "soak.failures",
   "value": 4,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Failed tasks, all clients. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "4"
  },
  {
   "id": "soak.claude_tasks",
   "value": 60,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Claude Code + Strim tasks run. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "60"
  },
  {
   "id": "soak.claude_passed",
   "value": 56,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Claude Code + Strim tasks passed (check + client exit 0). 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. Pass = deterministic check + client exit 0 within the cap; failures are listed in the summary file.",
   "dataset": "soak",
   "verify": "56"
  },
  {
   "id": "soak.codex_tasks",
   "value": 60,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Codex + Strim tasks run. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "60"
  },
  {
   "id": "soak.codex_passed",
   "value": 60,
   "unit": "tasks",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Codex + Strim tasks passed (check + client exit 0). 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. Pass = deterministic check + client exit 0 within the cap; failures are listed in the summary file.",
   "dataset": "soak",
   "verify": "60"
  },
  {
   "id": "soak.elapsed_min",
   "value": 30.3,
   "unit": "min",
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Wall-clock length of the run. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "30.3"
  },
  {
   "id": "soak.rss_min",
   "value": 8687.0,
   "unit": "MB",
   "dp": 0,
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Lowest server RSS sample. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "8687.0"
  },
  {
   "id": "soak.rss_max",
   "value": 8713.4,
   "unit": "MB",
   "dp": 0,
   "source_path": "bench/results/soak-20261027/run2/summary.txt",
   "method_note": "Highest server RSS sample. 30-minute real-client soak (earlier soak harness): claude -p and codex exec alternating on a toy repo, one concurrent pair every 6 tasks.",
   "caveat": "Earlier build and earlier harness (2026-10-27); not repeated on the current build or the current sandboxed harness. Single complete run of at least 30 minutes. ",
   "dataset": "soak",
   "verify": "8713.4"
  },
  {
   "id": "soak.rerun_a",
   "value": "1/4",
   "unit": "tasks passed",
   "source_path": "bench/results/soak-20261027/rerun-a/summary.txt",
   "method_note": "Claude Code tasks passed in the first rerun, before the harness allow-list fix (4 sequential claude -p tasks, not part of the soak).",
   "caveat": "Short rerun, not part of the 30-minute result.",
   "verify": "Tasks: 4 total, 1 passed"
  },
  {
   "id": "soak.rerun_b",
   "value": "4/4",
   "unit": "tasks passed",
   "source_path": "bench/results/soak-20261027/rerun-b/summary.txt",
   "method_note": "Claude Code tasks passed after the harness allow-list fix (4 sequential claude -p tasks, not part of the soak).",
   "caveat": "Short rerun, not part of the 30-minute result.",
   "verify": "Tasks: 4 total, 4 passed"
  },
  {
   "id": "mtp.depth_range",
   "value": "2-4",
   "unit": "tokens",
   "source_path": "docs/STATUS.md",
   "method_note": "Adaptive MTP draft depth range.",
   "caveat": "From the status document; fixed depth 2 is the opt-out (STRIM_MTP_FIXED=1).",
   "verify": "2–4"
  },
  {
   "id": "bg.block",
   "value": 48,
   "unit": "tokens",
   "source_path": "docs/BLOCKED-GDN-PREFILL-2026-10-26.md",
   "method_note": "Block size staged in threadgroup memory by the blocked GDN kernel.",
   "caveat": "",
   "verify": "48-token"
  },
  {
   "id": "cfg.chunk",
   "value": 1024,
   "unit": "tokens",
   "source_path": "bench/results/prefix-cache-exact-20261026/checkprefix-4096.log",
   "method_note": "Aligned prefill chunk / trie node size.",
   "caveat": "Printed as chunk=1024 by checkprefix.",
   "verify": "chunk=1024"
  },
  {
   "id": "ver.mlx_swift",
   "value": "MLX Swift 0.32.3 (Cmlx)",
   "unit": "",
   "source_path": "docs/PREFILL-SPIKE-2026-10-26.md",
   "method_note": "Version of MLX Swift.",
   "caveat": "",
   "verify": "0.32.3"
  },
  {
   "id": "ver.claude",
   "value": "Claude Code 2.1.310",
   "unit": "",
   "source_path": "bench/soak/README.md",
   "method_note": "Version of Claude Code (soak).",
   "caveat": "",
   "verify": "2.1.310"
  },
  {
   "id": "ver.codex",
   "value": "codex 0.164.0",
   "unit": "",
   "source_path": "bench/soak/README.md",
   "method_note": "Version of Codex (soak).",
   "caveat": "",
   "verify": "0.164.0"
  },
  {
   "id": "harness.tools",
   "value": 4,
   "unit": "tools",
   "source_path": "bench/results/mcp-smoke-20261027/summary.md",
   "method_note": "MCP tools/list on strim serve (strim_task, strim_job, strim_cancel, strim_ask).",
   "caveat": "Smoke run, one server, one session.",
   "dataset": "harness",
   "verify": "4 tools: strim_task"
  },
  {
   "id": "harness.rerun_runs",
   "value": 4,
   "unit": "runs",
   "source_path": "bench/results/mcp-clients-20261027/summary.md",
   "method_note": "Real Claude Code and Codex runs against the stock release build after the MCP fixes (2 Claude, 2 Codex).",
   "caveat": "n=1 per cell; trivial tasks (ask a question, fix a one-line bug).",
   "dataset": "harness",
   "verify_count": {
    "section": "Rerun after fixes",
    "contains": "|",
    "line_prefix": [
     "| claude_",
     "| codex_"
    ]
   }
  },
  {
   "id": "harness.rerun_pass",
   "value": 4,
   "unit": "runs passed",
   "source_path": "bench/results/mcp-clients-20261027/summary.md",
   "method_note": "Runs of the rerun table whose deterministic check passed (answer correct / git apply ok and tests pass, test files unchanged).",
   "caveat": "n=1 per cell; says the delegation path works, not that it saves tokens.",
   "dataset": "harness",
   "verify_count": {
    "section": "Rerun after fixes",
    "contains": "| PASS",
    "line_prefix": [
     "| claude_",
     "| codex_"
    ]
   }
  },
  {
   "id": "harness.rerun_n",
   "value": 1,
   "unit": "run per cell",
   "source_path": "bench/results/mcp-clients-20261027/summary.md",
   "method_note": "Sample size of each cell in the rerun table.",
   "caveat": "One run per client and prompt; parent output varies run to run.",
   "dataset": "harness",
   "verify": "n=1 each",
   "verify_section": "Rerun after fixes"
  },
  {
   "id": "harness.delegated_in",
   "value": "29.9K",
   "unit": "input tokens",
   "source_path": "docs/PLAN-PRODUCT-2026-10-27.md",
   "method_note": "Parent total input when a one-line fix was delegated through strim_task (3 turns).",
   "caveat": "Trivial task, n=1: no saving shown.",
   "dataset": "harness",
   "verify": "29.9K total input"
  },
  {
   "id": "harness.alone_in",
   "value": "25.5K",
   "unit": "input tokens",
   "source_path": "docs/PLAN-PRODUCT-2026-10-27.md",
   "method_note": "Parent total input when the same fix was done alone (5 turns).",
   "caveat": "Trivial task, n=1: no saving shown.",
   "dataset": "harness",
   "verify": "alone 25.5K"
  },
  {
   "id": "harness.agentbench_tasks",
   "value": 30,
   "unit": "tasks",
   "source_path": "docs/PLAN-PRODUCT-2026-10-27.md",
   "method_note": "Size of the real-task corpus the 3-arm agent benchmark will run on (not yet run).",
   "caveat": "Planned measurement; no result exists yet.",
   "dataset": "harness",
   "verify": "30 real tasks"
  },
  {
   "id": "ver.rapid",
   "value": "rapid-mlx 0.15.7",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Version of rapid-mlx in the final head-to-head.",
   "caveat": "MTP on, --max-num-seqs 4.",
   "verify": "rapid-mlx 0.15.7"
  },
  {
   "id": "ver.mlxvlm",
   "value": "mlx-vlm 0.7.10",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Version of mlx-vlm in the final head-to-head.",
   "caveat": "Server defaults, no MTP.",
   "verify": "mlx-vlm 0.7.10"
  },
  {
   "id": "ver.strim_h2h",
   "value": "strim 0.1.1 (ee9f9d8)",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim build in the final head-to-head.",
   "caveat": "Commits after this build were not benchmarked.",
   "verify": "ee9f9d8"
  },
  {
   "id": "ver.macos",
   "value": "27.0.1",
   "unit": "",
   "source_path": "bench/results/h2h-final-20261028/summary.md",
   "method_note": "macOS version of the benchmark machine.",
   "caveat": "",
   "verify": "macOS 27.0.1"
  },
  {
   "id": "ver.hw",
   "value": "Apple M5 Max, 64 GB",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Benchmark machine.",
   "caveat": "The only machine Strim has been tested on.",
   "computed": true
  },
  {
   "id": "cf.replay_before",
   "value": 52.3,
   "unit": "s",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Agent replay (8 turns, 4.1K to 40.8K tokens) summed prefill before the assistant-start prefix cuts: mean of 4 ABAB runs of the previous engine build.",
   "caveat": "Synthetic replay, same session as the other cache-footprint figures; absolute values move with machine load.",
   "verify": "(mean 52.3)"
  },
  {
   "id": "cf.replay_after_ab",
   "value": "49.6 / 49.9",
   "unit": "s",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Same replay with assistant-start cuts on by default (two runs, committed tree).",
   "caveat": "Two runs, not a mean; the final head-to-head measured 49.7 s for the same replay on the same build.",
   "verify": "replay 49.6 / 49.9 s"
  },
  {
   "id": "cf.cached_t2_before",
   "value": 3072,
   "unit": "tokens",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Cached prompt tokens at replay turn 2 before the assistant-start cuts.",
   "caveat": "Synthetic replay.",
   "verify": "(old 3072 8741"
  },
  {
   "id": "cf.cached_t2_after",
   "value": 4084,
   "unit": "tokens",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Cached prompt tokens at replay turn 2 with the assistant-start cuts (turn end becomes a trie node).",
   "caveat": "Synthetic replay.",
   "verify": "4084 9339"
  },
  {
   "id": "cf.fp_old_lo",
   "value": 31.1,
   "unit": "GB",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Lowest per-cell peak physical footprint before the cache and memory work (replay1 31128 MB).",
   "caveat": "Peak phys_footprint of the engine process, one run per cell.",
   "verify": "replay1 31128/27163"
  },
  {
   "id": "cf.fp_old_hi",
   "value": 41.9,
   "unit": "GB",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Highest per-cell peak footprint before (W5, 41894 MB).",
   "caveat": "Peak phys_footprint, one run per cell.",
   "verify": "w5 41894/32573"
  },
  {
   "id": "cf.fp_new_lo",
   "value": 27.0,
   "unit": "GB",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Lowest per-cell peak footprint after (replay2, 26983 MB).",
   "caveat": "Peak phys_footprint, one run per cell.",
   "verify": "replay2 36848/26983"
  },
  {
   "id": "cf.fp_new_hi",
   "value": 32.6,
   "unit": "GB",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Highest per-cell peak footprint after (W5, 32573 MB).",
   "caveat": "Peak phys_footprint, one run per cell.",
   "verify": "41894/32573"
  },
  {
   "id": "cf.cold_16k",
   "value": "17.2-17.3",
   "unit": "s",
   "source_path": "bench/results/cache-footprint-20261027/NOTES.md",
   "method_note": "Cold 16K TTFT in the old and the new build (no regression).",
   "caveat": "One session.",
   "verify": "16K 17.2-17.3 s"
  },
  {
   "id": "cb.c4_before",
   "value": 57.4,
   "unit": "tok/s",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Served aggregate decode at 4 concurrent 512-token requests before the batch-dimension GDN decode and conv kernels.",
   "caveat": "Measured under load (the plan states 2-7 load average), so the absolute values are lower than the quiet-machine head-to-head; compare within the pair only. Already included in the build used for the head-to-head.",
   "verify": "c=4 57.4 -> 63.3"
  },
  {
   "id": "cb.c4_after",
   "value": 63.3,
   "unit": "tok/s",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Same, with the batched kernels (default on).",
   "caveat": "Same load caveat; included in the head-to-head build, do not add it to the head-to-head figures.",
   "verify": "c=4 57.4 -> 63.3"
  },
  {
   "id": "cb.c8_before",
   "value": 66.2,
   "unit": "tok/s",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Served aggregate decode at 8 concurrent 512-token requests before the kernels.",
   "caveat": "Same load caveat.",
   "verify": "c=8 66.2 -> 73.1"
  },
  {
   "id": "cb.c8_after",
   "value": 73.1,
   "unit": "tok/s",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Same, with the batched kernels.",
   "caveat": "Same load caveat; included in the head-to-head build.",
   "verify": "c=8 66.2 -> 73.1"
  },
  {
   "id": "cb.step_before",
   "value": 49.5,
   "unit": "ms/step",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Batched decode step time at B=4, 512 context, before the batch-dimension kernels.",
   "caveat": "benchbatch microbenchmark.",
   "verify": "B=4 49.5 -> 42.3 ms/step"
  },
  {
   "id": "cb.step_after",
   "value": 42.3,
   "unit": "ms/step",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Batched decode step time at B=4 with the kernels.",
   "caveat": "benchbatch microbenchmark.",
   "verify": "B=4 49.5 -> 42.3 ms/step"
  },
  {
   "id": "cb.step_floor",
   "value": 34,
   "unit": "ms",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Plain single-stream decode step (weight-bandwidth floor).",
   "caveat": "From the plan document.",
   "verify": "B=1 plain step 34 ms"
  },
  {
   "id": "cb.agree_off",
   "value": "6/8",
   "unit": "prompts identical",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Batched decode without the kernels: prompts whose 96-token greedy output equals their own single-stream run (batch of 8).",
   "caveat": "8 fixed realistic prompts; each divergence is a single near-tie.",
   "verify": "| 6/8, 6/8 | 6/8, 6/8 |"
  },
  {
   "id": "cb.agree_on",
   "value": "8/8",
   "unit": "prompts identical",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Same with the batch kernels on: prompts identical to their own single-stream run at 96 tokens (7/8 at 128 tokens and 6K characters).",
   "caveat": "8 fixed realistic prompts; divergences are bf16 near-ties from the M>1 quantized matmul, not from the GDN kernels.",
   "verify": "| 8/8, 8/8 | 7/8, 7/8 |"
  },
  {
   "id": "cb.mtp_bound",
   "value": "1.09x",
   "unit": "x",
   "source_path": "docs/PLAN-CONCURRENCY-2026-10-27.md",
   "method_note": "Upper bound for MTP inside the batch (before draft cost): measured no-go.",
   "caveat": "Computed from measured verify cost and acceptance.",
   "verify": "<= 1.09x before draft"
  },
  {
   "id": "ad.max_inflight",
   "value": 8,
   "unit": "requests",
   "source_path": "docs/ADMISSION.md",
   "method_note": "Default total admission slots (STRIM_MAX_SESSIONS); background work is capped one below.",
   "caveat": "An admission cap, not the engine batch size; the engine admits by memory.",
   "verify": "STRIM_MAX_SESSIONS` (8)"
  },
  {
   "id": "sec.server_findings",
   "value": 14,
   "unit": "findings",
   "source_path": "docs/SECURITY-REVIEW-2026-10-28-server.md",
   "method_note": "Findings (F1 to F14) of the server HTTP surface review.",
   "caveat": "Read-only review plus an isolated arithmetic repro; no live server, no GPU.",
   "verify_count": {
    "contains": "### F",
    "line_prefix": [
     "### F"
    ]
   }
  },
  {
   "id": "sec.server_fixed",
   "value": 6,
   "unit": "findings",
   "source_path": "docs/SECURITY-REVIEW-2026-10-28-server.md",
   "method_note": "Server findings marked Fixed in the fix-status table (F1, F2, F3, F4, F6, F8).",
   "caveat": "A seventh (F7) is fixed partly; others are residual or were handled by the process-launch review.",
   "verify_count": {
    "contains": "| Fixed |",
    "line_prefix": [
     "| F"
    ]
   }
  },
  {
   "id": "sec.server_high",
   "value": 1,
   "unit": "findings",
   "source_path": "docs/SECURITY-REVIEW-2026-10-28-server.md",
   "method_note": "High-severity server findings (F1, integer overflow trap from a client-controlled token count), fixed.",
   "caveat": "No critical findings.",
   "verify": "High 1 (F1)"
  },
  {
   "id": "sec.exec_fixed",
   "value": 5,
   "unit": "findings",
   "source_path": "docs/SECURITY-REVIEW-2026-10-28-exec.md",
   "method_note": "Process-launch and model-influenced-code findings fixed (H1, M1 to M4).",
   "caveat": "The low and info items (L1 to L5, I1 to I3) were not addressed; residuals are listed per finding.",
   "verify_count": {
    "contains": "fixed:",
    "line_prefix": [
     "| H1",
     "| M"
    ]
   }
  },
  {
   "id": "qt.mem_gib",
   "value": "56.0",
   "unit": "GiB",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "Metal recommendedMaxWorkingSetSize on the 64 GB test machine (87.5% of RAM).",
   "caveat": "Measured on one machine and one macOS version; other tiers are not measured.",
   "verify": "**56.0 GiB = 87.5 % of RAM**"
  },
  {
   "id": "qt.w4",
   "value": "14.09",
   "unit": "GiB",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "Computed text-only weight size of a 4-bit affine g64 quantization of the 27B backbone.",
   "caveat": "Computed from the config, not measured; the checkpoint on disk also carries a vision tower.",
   "verify": "4-bit 14.09"
  },
  {
   "id": "qt.f4",
   "value": "17.0",
   "unit": "GiB",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "Computed fixed engine overhead F for 4-bit g64 (weights, MTP head, one GDN state, MLX cache and prefill transient).",
   "caveat": "Computed with assumptions, not measured on other tiers.",
   "verify": "4-bit g64 (F 17.0)"
  },
  {
   "id": "qt.t8",
   "value": "no",
   "unit": "",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "8 GB tier: even a 2-bit 27B exceeds the GPU budget.",
   "caveat": "Computed.",
   "verify": "**8 GB: does not fit.**"
  },
  {
   "id": "qt.t16",
   "value": "no",
   "unit": "",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "16 GB tier: does not fit in practice.",
   "caveat": "Computed; no 4-bit configuration fits.",
   "verify": "**16 GB: does not fit in practice.**"
  },
  {
   "id": "qt.t24",
   "value": "15K/29K / 23K/45K",
   "unit": "tokens",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "24 GB tier, 4-bit g64: context estimate (bf16 KV / q8 KV) at the 75% and 87.5% GPU budget.",
   "caveat": "Computed, not measured; uses an assumed 5.5 GiB for macOS and apps.",
   "verify": "15K/29K / 23K/45K"
  },
  {
   "id": "qt.t32",
   "value": "114K/214K (low) 154K/291K (high)",
   "unit": "tokens",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "32 GB tier, 4-bit g64: context estimate (bf16 KV / q8 KV) at the low and high GPU budget.",
   "caveat": "Computed, not measured.",
   "verify": "114K/214K (low) 154K/291K (high)"
  },
  {
   "id": "qt.t36",
   "value": "163K/307K - 220K+",
   "unit": "tokens",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "36 GB tier, 4-bit g64: context estimate (bf16 KV / q8 KV).",
   "caveat": "Computed, not measured.",
   "verify": "163K/307K - 220K+"
  },
  {
   "id": "qt.t48",
   "value": "310K+",
   "unit": "tokens",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "48 GB tier, 4-bit g64: context estimate (bf16 KV).",
   "caveat": "Computed, not measured.",
   "verify": "| 310K+ |"
  },
  {
   "id": "qt.t64",
   "value": "full",
   "unit": "",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "64 GB tier, 4-bit g64: the full 262K native context fits (single sequence, no prefix cache).",
   "caveat": "Computed; the 64 GB tier is also the only measured one.",
   "verify": "| 64 GB | full | full | full |"
  },
  {
   "id": "ab.runs",
   "value": 120,
   "unit": "runs",
   "source_path": "bench/results/agentbench-20261028/report.md",
   "method_note": "Runs of the free arms (H harness, C Claude Code client on the local model) in the agent benchmark of 2026-10-28.",
   "caveat": "Earlier 30-task corpus: some tasks had an unusable test environment inside the sandboxes, so those results measure the environment, not the model. No pass rate from this run is quoted.",
   "verify": "agentbench report: 120 runs"
  },
  {
   "id": "ab.corpus",
   "value": 17,
   "unit": "tasks",
   "source_path": "docs/RESEARCH-QUANT-TIERS-2026-10-27.md",
   "method_note": "Tasks in the current agent task corpus (12 test-fix, 3 mechanical-edit, 2 repo-qa; Swift and Go).",
   "caveat": "Not yet re-run with the agent benchmark; the research note says 17 tasks are too few for fine decisions.",
   "verify": "the 17 tasks in `bench/agenttasks`"
  },
  {
   "id": "final.c4k_pf.strim",
   "value": 994.5,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K prefill, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "strim",
    "cell": "994.5"
   }
  },
  {
   "id": "final.c4k_pf.rapid-mlx",
   "value": 996.8,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K prefill, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "rapid-mlx",
    "cell": "996.8"
   }
  },
  {
   "id": "final.c4k_pf.mlx-vlm",
   "value": 1017.0,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K prefill, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 0,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "mlx-vlm",
    "cell": "1017"
   }
  },
  {
   "id": "final.c4k_pf.mlx-lm",
   "value": 949.1,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K prefill, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "mlx-lm",
    "cell": "949.1"
   }
  },
  {
   "id": "final.c4k_pf.omlx",
   "value": 896.7,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K prefill, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "omlx",
    "cell": "896.7"
   }
  },
  {
   "id": "final.c4k_pf.ratio.rapid-mlx",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K prefill, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "strim/rapid-mlx",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.c4k_pf.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K prefill against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.c4k_pf.ratio.mlx-vlm",
   "value": 0.98,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K prefill, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "strim/mlx-vlm",
    "cell": "0.98x"
   }
  },
  {
   "id": "final.c4k_pf.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K prefill against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c4k_pf.ratio.mlx-lm",
   "value": 1.05,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K prefill, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "strim/mlx-lm",
    "cell": "1.05x"
   }
  },
  {
   "id": "final.c4k_pf.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K prefill against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c4k_pf.ratio.omlx",
   "value": 1.11,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K prefill, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "strim/omlx",
    "cell": "1.11x"
   }
  },
  {
   "id": "final.c4k_pf.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K prefill against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K prefill",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c4k_tt.strim",
   "value": 4.13,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "strim",
    "cell": "4.13"
   }
  },
  {
   "id": "final.c4k_tt.rapid-mlx",
   "value": 4.12,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "rapid-mlx",
    "cell": "4.12"
   }
  },
  {
   "id": "final.c4k_tt.mlx-vlm",
   "value": 4.039,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "mlx-vlm",
    "cell": "4.039"
   }
  },
  {
   "id": "final.c4k_tt.mlx-lm",
   "value": 4.327,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "mlx-lm",
    "cell": "4.327"
   }
  },
  {
   "id": "final.c4k_tt.omlx",
   "value": 4.58,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 4K TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "omlx",
    "cell": "4.58"
   }
  },
  {
   "id": "final.c4k_tt.ratio.rapid-mlx",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "strim/rapid-mlx",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.c4k_tt.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.c4k_tt.ratio.mlx-vlm",
   "value": 1.02,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "strim/mlx-vlm",
    "cell": "1.02x"
   }
  },
  {
   "id": "final.c4k_tt.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c4k_tt.ratio.mlx-lm",
   "value": 0.95,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.95x"
   }
  },
  {
   "id": "final.c4k_tt.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c4k_tt.ratio.omlx",
   "value": 0.9,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 4K TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "strim/omlx",
    "cell": "0.90x"
   }
  },
  {
   "id": "final.c4k_tt.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 4K TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 4K TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d4k.strim",
   "value": 75.34,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 4K ctx, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "strim",
    "cell": "75.34"
   }
  },
  {
   "id": "final.d4k.rapid-mlx",
   "value": 50.12,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 4K ctx, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "rapid-mlx",
    "cell": "50.12"
   }
  },
  {
   "id": "final.d4k.mlx-vlm",
   "value": 32.69,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 4K ctx, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "mlx-vlm",
    "cell": "32.69"
   }
  },
  {
   "id": "final.d4k.mlx-lm",
   "value": 30.11,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 4K ctx, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "mlx-lm",
    "cell": "30.11"
   }
  },
  {
   "id": "final.d4k.omlx",
   "value": 33.53,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 4K ctx, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "omlx",
    "cell": "33.53"
   }
  },
  {
   "id": "final.d4k.ratio.rapid-mlx",
   "value": 1.5,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 4K ctx, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "strim/rapid-mlx",
    "cell": "1.50x"
   }
  },
  {
   "id": "final.d4k.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 4K ctx against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d4k.ratio.mlx-vlm",
   "value": 2.3,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 4K ctx, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "strim/mlx-vlm",
    "cell": "2.30x"
   }
  },
  {
   "id": "final.d4k.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 4K ctx against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d4k.ratio.mlx-lm",
   "value": 2.5,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 4K ctx, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "strim/mlx-lm",
    "cell": "2.50x"
   }
  },
  {
   "id": "final.d4k.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 4K ctx against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d4k.ratio.omlx",
   "value": 2.25,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 4K ctx, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "strim/omlx",
    "cell": "2.25x"
   }
  },
  {
   "id": "final.d4k.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 4K ctx against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 4K ctx",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c16k_pf.strim",
   "value": 963.1,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K prefill, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "strim",
    "cell": "963.1"
   }
  },
  {
   "id": "final.c16k_pf.rapid-mlx",
   "value": 972.4,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K prefill, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "rapid-mlx",
    "cell": "972.4"
   }
  },
  {
   "id": "final.c16k_pf.mlx-vlm",
   "value": 988.0,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K prefill, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 0,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "mlx-vlm",
    "cell": "988"
   }
  },
  {
   "id": "final.c16k_pf.mlx-lm",
   "value": 919.8,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K prefill, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "mlx-lm",
    "cell": "919.8"
   }
  },
  {
   "id": "final.c16k_pf.omlx",
   "value": 891.9,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K prefill, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "omlx",
    "cell": "891.9"
   }
  },
  {
   "id": "final.c16k_pf.ratio.rapid-mlx",
   "value": 0.99,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K prefill, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "strim/rapid-mlx",
    "cell": "0.99x"
   }
  },
  {
   "id": "final.c16k_pf.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K prefill against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.c16k_pf.ratio.mlx-vlm",
   "value": 0.97,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K prefill, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "strim/mlx-vlm",
    "cell": "0.97x"
   }
  },
  {
   "id": "final.c16k_pf.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K prefill against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c16k_pf.ratio.mlx-lm",
   "value": 1.05,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K prefill, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "strim/mlx-lm",
    "cell": "1.05x"
   }
  },
  {
   "id": "final.c16k_pf.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K prefill against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c16k_pf.ratio.omlx",
   "value": 1.08,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K prefill, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "strim/omlx",
    "cell": "1.08x"
   }
  },
  {
   "id": "final.c16k_pf.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K prefill against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K prefill",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c16k_tt.strim",
   "value": 17.02,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "strim",
    "cell": "17.02"
   }
  },
  {
   "id": "final.c16k_tt.rapid-mlx",
   "value": 16.86,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "rapid-mlx",
    "cell": "16.86"
   }
  },
  {
   "id": "final.c16k_tt.mlx-vlm",
   "value": 16.59,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "mlx-vlm",
    "cell": "16.59"
   }
  },
  {
   "id": "final.c16k_tt.mlx-lm",
   "value": 17.82,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "mlx-lm",
    "cell": "17.82"
   }
  },
  {
   "id": "final.c16k_tt.omlx",
   "value": 18.38,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 16K TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "omlx",
    "cell": "18.38"
   }
  },
  {
   "id": "final.c16k_tt.ratio.rapid-mlx",
   "value": 1.01,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "strim/rapid-mlx",
    "cell": "1.01x"
   }
  },
  {
   "id": "final.c16k_tt.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.c16k_tt.ratio.mlx-vlm",
   "value": 1.03,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "strim/mlx-vlm",
    "cell": "1.03x"
   }
  },
  {
   "id": "final.c16k_tt.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c16k_tt.ratio.mlx-lm",
   "value": 0.96,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.96x"
   }
  },
  {
   "id": "final.c16k_tt.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c16k_tt.ratio.omlx",
   "value": 0.93,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 16K TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "strim/omlx",
    "cell": "0.93x"
   }
  },
  {
   "id": "final.c16k_tt.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 16K TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 16K TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d16k.strim",
   "value": 65.18,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 16K ctx, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "strim",
    "cell": "65.18"
   }
  },
  {
   "id": "final.d16k.rapid-mlx",
   "value": 45.75,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 16K ctx, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "rapid-mlx",
    "cell": "45.75"
   }
  },
  {
   "id": "final.d16k.mlx-vlm",
   "value": 30.93,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 16K ctx, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "mlx-vlm",
    "cell": "30.93"
   }
  },
  {
   "id": "final.d16k.mlx-lm",
   "value": 28.68,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 16K ctx, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "mlx-lm",
    "cell": "28.68"
   }
  },
  {
   "id": "final.d16k.omlx",
   "value": 31.42,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 256 tok @ 16K ctx, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "omlx",
    "cell": "31.42"
   }
  },
  {
   "id": "final.d16k.ratio.rapid-mlx",
   "value": 1.42,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 16K ctx, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "strim/rapid-mlx",
    "cell": "1.42x"
   }
  },
  {
   "id": "final.d16k.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 16K ctx against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d16k.ratio.mlx-vlm",
   "value": 2.11,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 16K ctx, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "strim/mlx-vlm",
    "cell": "2.11x"
   }
  },
  {
   "id": "final.d16k.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 16K ctx against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d16k.ratio.mlx-lm",
   "value": 2.27,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 16K ctx, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "strim/mlx-lm",
    "cell": "2.27x"
   }
  },
  {
   "id": "final.d16k.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 16K ctx against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d16k.ratio.omlx",
   "value": 2.07,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 256 tok @ 16K ctx, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "strim/omlx",
    "cell": "2.07x"
   }
  },
  {
   "id": "final.d16k.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 256 tok @ 16K ctx against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 256 tok @ 16K ctx",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c32k_pf.strim",
   "value": 884.6,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K prefill, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "strim",
    "cell": "884.6"
   }
  },
  {
   "id": "final.c32k_pf.rapid-mlx",
   "value": 910.0,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K prefill, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 0,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "rapid-mlx",
    "cell": "910"
   }
  },
  {
   "id": "final.c32k_pf.mlx-vlm",
   "value": 923.2,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K prefill, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "mlx-vlm",
    "cell": "923.2"
   }
  },
  {
   "id": "final.c32k_pf.mlx-lm",
   "value": 850.8,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K prefill, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "mlx-lm",
    "cell": "850.8"
   }
  },
  {
   "id": "final.c32k_pf.omlx",
   "value": 823.7,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K prefill, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "omlx",
    "cell": "823.7"
   }
  },
  {
   "id": "final.c32k_pf.ratio.rapid-mlx",
   "value": 0.97,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K prefill, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "strim/rapid-mlx",
    "cell": "0.97x"
   }
  },
  {
   "id": "final.c32k_pf.verdict.rapid-mlx",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K prefill against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "vs rapid-mlx",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c32k_pf.ratio.mlx-vlm",
   "value": 0.96,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K prefill, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "strim/mlx-vlm",
    "cell": "0.96x"
   }
  },
  {
   "id": "final.c32k_pf.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K prefill against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c32k_pf.ratio.mlx-lm",
   "value": 1.04,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K prefill, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "strim/mlx-lm",
    "cell": "1.04x"
   }
  },
  {
   "id": "final.c32k_pf.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K prefill against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c32k_pf.ratio.omlx",
   "value": 1.07,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K prefill, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "strim/omlx",
    "cell": "1.07x"
   }
  },
  {
   "id": "final.c32k_pf.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K prefill against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K prefill",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c32k_tt.strim",
   "value": 37.06,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "strim",
    "cell": "37.06"
   }
  },
  {
   "id": "final.c32k_tt.rapid-mlx",
   "value": 36.02,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "rapid-mlx",
    "cell": "36.02"
   }
  },
  {
   "id": "final.c32k_tt.mlx-vlm",
   "value": 35.51,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "mlx-vlm",
    "cell": "35.51"
   }
  },
  {
   "id": "final.c32k_tt.mlx-lm",
   "value": 38.53,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "mlx-lm",
    "cell": "38.53"
   }
  },
  {
   "id": "final.c32k_tt.omlx",
   "value": 39.79,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): cold 32K TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "omlx",
    "cell": "39.79"
   }
  },
  {
   "id": "final.c32k_tt.ratio.rapid-mlx",
   "value": 1.03,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "strim/rapid-mlx",
    "cell": "1.03x"
   }
  },
  {
   "id": "final.c32k_tt.verdict.rapid-mlx",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "vs rapid-mlx",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c32k_tt.ratio.mlx-vlm",
   "value": 1.04,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "strim/mlx-vlm",
    "cell": "1.04x"
   }
  },
  {
   "id": "final.c32k_tt.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.c32k_tt.ratio.mlx-lm",
   "value": 0.96,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.96x"
   }
  },
  {
   "id": "final.c32k_tt.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.c32k_tt.ratio.omlx",
   "value": 0.93,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: cold 32K TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "strim/omlx",
    "cell": "0.93x"
   }
  },
  {
   "id": "final.c32k_tt.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for cold 32K TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "cold 32K TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d32k.strim",
   "value": 52.17,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 128 tok @ 32K ctx, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "strim",
    "cell": "52.17"
   }
  },
  {
   "id": "final.d32k.rapid-mlx",
   "value": 42.17,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 128 tok @ 32K ctx, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "rapid-mlx",
    "cell": "42.17"
   }
  },
  {
   "id": "final.d32k.mlx-vlm",
   "value": 29.35,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 128 tok @ 32K ctx, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "mlx-vlm",
    "cell": "29.35"
   }
  },
  {
   "id": "final.d32k.mlx-lm",
   "value": 27.62,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 128 tok @ 32K ctx, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "mlx-lm",
    "cell": "27.62"
   }
  },
  {
   "id": "final.d32k.omlx",
   "value": 29.9,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): decode 128 tok @ 32K ctx, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "omlx",
    "cell": "29.9"
   }
  },
  {
   "id": "final.d32k.ratio.rapid-mlx",
   "value": 1.24,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 128 tok @ 32K ctx, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "strim/rapid-mlx",
    "cell": "1.24x"
   }
  },
  {
   "id": "final.d32k.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 128 tok @ 32K ctx against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d32k.ratio.mlx-vlm",
   "value": 1.78,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 128 tok @ 32K ctx, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "strim/mlx-vlm",
    "cell": "1.78x"
   }
  },
  {
   "id": "final.d32k.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 128 tok @ 32K ctx against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d32k.ratio.mlx-lm",
   "value": 1.89,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 128 tok @ 32K ctx, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "strim/mlx-lm",
    "cell": "1.89x"
   }
  },
  {
   "id": "final.d32k.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 128 tok @ 32K ctx against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.d32k.ratio.omlx",
   "value": 1.74,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: decode 128 tok @ 32K ctx, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "strim/omlx",
    "cell": "1.74x"
   }
  },
  {
   "id": "final.d32k.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for decode 128 tok @ 32K ctx against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "decode 128 tok @ 32K ctx",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c4.strim",
   "value": 69.22,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts aggregate, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "strim",
    "cell": "69.22"
   }
  },
  {
   "id": "final.a512c4.rapid-mlx",
   "value": 71.69,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts aggregate, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "rapid-mlx",
    "cell": "71.69"
   }
  },
  {
   "id": "final.a512c4.mlx-vlm",
   "value": 67.29,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts aggregate, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "mlx-vlm",
    "cell": "67.29"
   }
  },
  {
   "id": "final.a512c4.mlx-lm",
   "value": 66.31,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts aggregate, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "mlx-lm",
    "cell": "66.31"
   }
  },
  {
   "id": "final.a512c4.omlx",
   "value": 54.61,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts aggregate, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "omlx",
    "cell": "54.61"
   }
  },
  {
   "id": "final.a512c4.ratio.rapid-mlx",
   "value": 0.97,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts aggregate, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "strim/rapid-mlx",
    "cell": "0.97x"
   }
  },
  {
   "id": "final.a512c4.verdict.rapid-mlx",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts aggregate against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "vs rapid-mlx",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.a512c4.ratio.mlx-vlm",
   "value": 1.03,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts aggregate, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "strim/mlx-vlm",
    "cell": "1.03x"
   }
  },
  {
   "id": "final.a512c4.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts aggregate against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c4.ratio.mlx-lm",
   "value": 1.04,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts aggregate, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "strim/mlx-lm",
    "cell": "1.04x"
   }
  },
  {
   "id": "final.a512c4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts aggregate against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c4.ratio.omlx",
   "value": 1.27,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts aggregate, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "strim/omlx",
    "cell": "1.27x"
   }
  },
  {
   "id": "final.a512c4.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts aggregate against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts aggregate",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c4.strim",
   "value": 1.493,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "strim",
    "cell": "1.493"
   }
  },
  {
   "id": "final.p50_512c4.rapid-mlx",
   "value": 2.339,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "rapid-mlx",
    "cell": "2.339"
   }
  },
  {
   "id": "final.p50_512c4.mlx-vlm",
   "value": 2.305,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "mlx-vlm",
    "cell": "2.305"
   }
  },
  {
   "id": "final.p50_512c4.mlx-lm",
   "value": 2.467,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "mlx-lm",
    "cell": "2.467"
   }
  },
  {
   "id": "final.p50_512c4.omlx",
   "value": 3.049,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p50 TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "omlx",
    "cell": "3.049"
   }
  },
  {
   "id": "final.p50_512c4.ratio.rapid-mlx",
   "value": 0.64,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.64x"
   }
  },
  {
   "id": "final.p50_512c4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c4.ratio.mlx-vlm",
   "value": 0.65,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.65x"
   }
  },
  {
   "id": "final.p50_512c4.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c4.ratio.mlx-lm",
   "value": 0.61,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.61x"
   }
  },
  {
   "id": "final.p50_512c4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c4.ratio.omlx",
   "value": 0.49,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p50 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "strim/omlx",
    "cell": "0.49x"
   }
  },
  {
   "id": "final.p50_512c4.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p50 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_512c4.strim",
   "value": 2.292,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "strim",
    "cell": "2.292"
   }
  },
  {
   "id": "final.p95_512c4.rapid-mlx",
   "value": 2.339,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "rapid-mlx",
    "cell": "2.339"
   }
  },
  {
   "id": "final.p95_512c4.mlx-vlm",
   "value": 2.305,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "mlx-vlm",
    "cell": "2.305"
   }
  },
  {
   "id": "final.p95_512c4.mlx-lm",
   "value": 2.469,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "mlx-lm",
    "cell": "2.469"
   }
  },
  {
   "id": "final.p95_512c4.omlx",
   "value": 3.198,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 512-tok prompts p95 TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "omlx",
    "cell": "3.198"
   }
  },
  {
   "id": "final.p95_512c4.ratio.rapid-mlx",
   "value": 0.98,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.98x"
   }
  },
  {
   "id": "final.p95_512c4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_512c4.ratio.mlx-vlm",
   "value": 0.99,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.99x"
   }
  },
  {
   "id": "final.p95_512c4.verdict.mlx-vlm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.p95_512c4.ratio.mlx-lm",
   "value": 0.93,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.93x"
   }
  },
  {
   "id": "final.p95_512c4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_512c4.ratio.omlx",
   "value": 0.72,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 512-tok prompts p95 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "strim/omlx",
    "cell": "0.72x"
   }
  },
  {
   "id": "final.p95_512c4.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 512-tok prompts p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 512-tok prompts p95 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c8.strim",
   "value": 76.37,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts aggregate, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "strim",
    "cell": "76.37"
   }
  },
  {
   "id": "final.a512c8.rapid-mlx",
   "value": 71.76,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts aggregate, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "rapid-mlx",
    "cell": "71.76"
   }
  },
  {
   "id": "final.a512c8.mlx-vlm",
   "value": 74.65,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts aggregate, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "mlx-vlm",
    "cell": "74.65"
   }
  },
  {
   "id": "final.a512c8.mlx-lm",
   "value": 67.9,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts aggregate, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "mlx-lm",
    "cell": "67.9"
   }
  },
  {
   "id": "final.a512c8.omlx",
   "value": 58.74,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts aggregate, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "omlx",
    "cell": "58.74"
   }
  },
  {
   "id": "final.a512c8.ratio.rapid-mlx",
   "value": 1.06,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts aggregate, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "strim/rapid-mlx",
    "cell": "1.06x"
   }
  },
  {
   "id": "final.a512c8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts aggregate against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c8.ratio.mlx-vlm",
   "value": 1.02,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts aggregate, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "strim/mlx-vlm",
    "cell": "1.02x"
   }
  },
  {
   "id": "final.a512c8.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts aggregate against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c8.ratio.mlx-lm",
   "value": 1.12,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts aggregate, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "strim/mlx-lm",
    "cell": "1.12x"
   }
  },
  {
   "id": "final.a512c8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts aggregate against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a512c8.ratio.omlx",
   "value": 1.3,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts aggregate, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "strim/omlx",
    "cell": "1.30x"
   }
  },
  {
   "id": "final.a512c8.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts aggregate against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts aggregate",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c8.strim",
   "value": 2.683,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "strim",
    "cell": "2.683"
   }
  },
  {
   "id": "final.p50_512c8.rapid-mlx",
   "value": 5.889,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "rapid-mlx",
    "cell": "5.889"
   }
  },
  {
   "id": "final.p50_512c8.mlx-vlm",
   "value": 4.34,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "mlx-vlm",
    "cell": "4.34"
   }
  },
  {
   "id": "final.p50_512c8.mlx-lm",
   "value": 6.064,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "mlx-lm",
    "cell": "6.064"
   }
  },
  {
   "id": "final.p50_512c8.omlx",
   "value": 5.726,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p50 TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "omlx",
    "cell": "5.726"
   }
  },
  {
   "id": "final.p50_512c8.ratio.rapid-mlx",
   "value": 0.46,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.46x"
   }
  },
  {
   "id": "final.p50_512c8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c8.ratio.mlx-vlm",
   "value": 0.62,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.62x"
   }
  },
  {
   "id": "final.p50_512c8.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c8.ratio.mlx-lm",
   "value": 0.44,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.44x"
   }
  },
  {
   "id": "final.p50_512c8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_512c8.ratio.omlx",
   "value": 0.47,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p50 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "strim/omlx",
    "cell": "0.47x"
   }
  },
  {
   "id": "final.p50_512c8.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p50 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_512c8.strim",
   "value": 4.61,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "strim",
    "cell": "4.61"
   }
  },
  {
   "id": "final.p95_512c8.rapid-mlx",
   "value": 9.466,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "rapid-mlx",
    "cell": "9.466"
   }
  },
  {
   "id": "final.p95_512c8.mlx-vlm",
   "value": 4.341,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "mlx-vlm",
    "cell": "4.341"
   }
  },
  {
   "id": "final.p95_512c8.mlx-lm",
   "value": 9.837,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "mlx-lm",
    "cell": "9.837"
   }
  },
  {
   "id": "final.p95_512c8.omlx",
   "value": 6.304,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 512-tok prompts p95 TTFT, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "omlx",
    "cell": "6.304"
   }
  },
  {
   "id": "final.p95_512c8.ratio.rapid-mlx",
   "value": 0.49,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.49x"
   }
  },
  {
   "id": "final.p95_512c8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_512c8.ratio.mlx-vlm",
   "value": 1.06,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "1.06x"
   }
  },
  {
   "id": "final.p95_512c8.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.p95_512c8.ratio.mlx-lm",
   "value": 0.47,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.47x"
   }
  },
  {
   "id": "final.p95_512c8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_512c8.ratio.omlx",
   "value": 0.73,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 512-tok prompts p95 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "strim/omlx",
    "cell": "0.73x"
   }
  },
  {
   "id": "final.p95_512c8.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 512-tok prompts p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 512-tok prompts p95 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aagc4.strim",
   "value": 28.73,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts (~2.8K tok) aggregate, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "strim",
    "cell": "28.73"
   }
  },
  {
   "id": "final.aagc4.rapid-mlx",
   "value": 27.73,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts (~2.8K tok) aggregate, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "rapid-mlx",
    "cell": "27.73"
   }
  },
  {
   "id": "final.aagc4.mlx-vlm",
   "value": 28.57,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts (~2.8K tok) aggregate, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "mlx-vlm",
    "cell": "28.57"
   }
  },
  {
   "id": "final.aagc4.mlx-lm",
   "value": 27.22,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts (~2.8K tok) aggregate, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "mlx-lm",
    "cell": "27.22"
   }
  },
  {
   "id": "final.aagc4.omlx",
   "value": 25.28,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts (~2.8K tok) aggregate, omlx. Median of 1 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "1",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "omlx",
    "cell": "25.28"
   }
  },
  {
   "id": "final.aagc4.ratio.rapid-mlx",
   "value": 1.04,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts (~2.8K tok) aggregate, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "strim/rapid-mlx",
    "cell": "1.04x"
   }
  },
  {
   "id": "final.aagc4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts (~2.8K tok) aggregate against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aagc4.ratio.mlx-vlm",
   "value": 1.01,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts (~2.8K tok) aggregate, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "strim/mlx-vlm",
    "cell": "1.01x"
   }
  },
  {
   "id": "final.aagc4.verdict.mlx-vlm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts (~2.8K tok) aggregate against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "vs mlx-vlm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.aagc4.ratio.mlx-lm",
   "value": 1.06,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts (~2.8K tok) aggregate, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "strim/mlx-lm",
    "cell": "1.06x"
   }
  },
  {
   "id": "final.aagc4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts (~2.8K tok) aggregate against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aagc4.ratio.omlx",
   "value": 1.14,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts (~2.8K tok) aggregate, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "strim/omlx",
    "cell": "1.14x"
   }
  },
  {
   "id": "final.aagc4.verdict.omlx",
   "value": "NC",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts (~2.8K tok) aggregate against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Not claimable: oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog). The summary.md verdict is withheld here because the cell is not claim_eligible.",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts (~2.8K tok) aggregate",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc4.strim",
   "value": 7.678,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "strim",
    "cell": "7.678"
   }
  },
  {
   "id": "final.p50_agc4.rapid-mlx",
   "value": 13.44,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "rapid-mlx",
    "cell": "13.44"
   }
  },
  {
   "id": "final.p50_agc4.mlx-vlm",
   "value": 12.43,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "mlx-vlm",
    "cell": "12.43"
   }
  },
  {
   "id": "final.p50_agc4.mlx-lm",
   "value": 13.47,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "mlx-lm",
    "cell": "13.47"
   }
  },
  {
   "id": "final.p50_agc4.omlx",
   "value": 13.32,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p50 TTFT, omlx. Median of 1 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "1",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "omlx",
    "cell": "13.32"
   }
  },
  {
   "id": "final.p50_agc4.ratio.rapid-mlx",
   "value": 0.57,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.57x"
   }
  },
  {
   "id": "final.p50_agc4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc4.ratio.mlx-vlm",
   "value": 0.62,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.62x"
   }
  },
  {
   "id": "final.p50_agc4.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc4.ratio.mlx-lm",
   "value": 0.57,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.57x"
   }
  },
  {
   "id": "final.p50_agc4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc4.ratio.omlx",
   "value": 0.58,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p50 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "strim/omlx",
    "cell": "0.58x"
   }
  },
  {
   "id": "final.p50_agc4.verdict.omlx",
   "value": "NC",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Not claimable: oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog). The summary.md verdict is withheld here because the cell is not claim_eligible.",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p50 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc4.strim",
   "value": 12.07,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "strim",
    "cell": "12.07"
   }
  },
  {
   "id": "final.p95_agc4.rapid-mlx",
   "value": 13.44,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "rapid-mlx",
    "cell": "13.44"
   }
  },
  {
   "id": "final.p95_agc4.mlx-vlm",
   "value": 12.43,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "mlx-vlm",
    "cell": "12.43"
   }
  },
  {
   "id": "final.p95_agc4.mlx-lm",
   "value": 13.47,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "mlx-lm",
    "cell": "13.47"
   }
  },
  {
   "id": "final.p95_agc4.omlx",
   "value": 13.47,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 agent prompts p95 TTFT, omlx. Median of 1 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "1",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "omlx",
    "cell": "13.47"
   }
  },
  {
   "id": "final.p95_agc4.ratio.rapid-mlx",
   "value": 0.9,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.90x"
   }
  },
  {
   "id": "final.p95_agc4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc4.ratio.mlx-vlm",
   "value": 0.97,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.97x"
   }
  },
  {
   "id": "final.p95_agc4.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc4.ratio.mlx-lm",
   "value": 0.9,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.90x"
   }
  },
  {
   "id": "final.p95_agc4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc4.ratio.omlx",
   "value": 0.9,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 agent prompts p95 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "strim/omlx",
    "cell": "0.90x"
   }
  },
  {
   "id": "final.p95_agc4.verdict.omlx",
   "value": "NC",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 agent prompts p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Not claimable: oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog). The summary.md verdict is withheld here because the cell is not claim_eligible.",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 agent prompts p95 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aagc8.strim",
   "value": 31.19,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts aggregate, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "strim",
    "cell": "31.19"
   }
  },
  {
   "id": "final.aagc8.rapid-mlx",
   "value": 28.47,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts aggregate, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "rapid-mlx",
    "cell": "28.47"
   }
  },
  {
   "id": "final.aagc8.mlx-vlm",
   "value": 31.18,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts aggregate, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "mlx-vlm",
    "cell": "31.18"
   }
  },
  {
   "id": "final.aagc8.mlx-lm",
   "value": 27.63,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts aggregate, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "mlx-lm",
    "cell": "27.63"
   }
  },
  {
   "id": "final.aagc8.omlx",
   "value": 26.89,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts aggregate, omlx. Median of 1 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "1",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "omlx",
    "cell": "26.89"
   }
  },
  {
   "id": "final.aagc8.ratio.rapid-mlx",
   "value": 1.1,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts aggregate, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "strim/rapid-mlx",
    "cell": "1.10x"
   }
  },
  {
   "id": "final.aagc8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts aggregate against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aagc8.ratio.mlx-vlm",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts aggregate, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "strim/mlx-vlm",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.aagc8.verdict.mlx-vlm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts aggregate against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "vs mlx-vlm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.aagc8.ratio.mlx-lm",
   "value": 1.13,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts aggregate, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "strim/mlx-lm",
    "cell": "1.13x"
   }
  },
  {
   "id": "final.aagc8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts aggregate against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aagc8.ratio.omlx",
   "value": 1.16,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts aggregate, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "strim/omlx",
    "cell": "1.16x"
   }
  },
  {
   "id": "final.aagc8.verdict.omlx",
   "value": "NC",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts aggregate against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Not claimable: oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog). The summary.md verdict is withheld here because the cell is not claim_eligible.",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts aggregate",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc8.strim",
   "value": 13.34,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "strim",
    "cell": "13.34"
   }
  },
  {
   "id": "final.p50_agc8.rapid-mlx",
   "value": 22.1,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "rapid-mlx",
    "cell": "22.1"
   }
  },
  {
   "id": "final.p50_agc8.mlx-vlm",
   "value": 23.1,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "mlx-vlm",
    "cell": "23.1"
   }
  },
  {
   "id": "final.p50_agc8.mlx-lm",
   "value": 22.61,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "mlx-lm",
    "cell": "22.61"
   }
  },
  {
   "id": "final.p50_agc8.omlx",
   "value": 25.08,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p50 TTFT, omlx. Median of 1 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "1",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "omlx",
    "cell": "25.08"
   }
  },
  {
   "id": "final.p50_agc8.ratio.rapid-mlx",
   "value": 0.6,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.60x"
   }
  },
  {
   "id": "final.p50_agc8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc8.ratio.mlx-vlm",
   "value": 0.58,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.58x"
   }
  },
  {
   "id": "final.p50_agc8.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc8.ratio.mlx-lm",
   "value": 0.59,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.59x"
   }
  },
  {
   "id": "final.p50_agc8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_agc8.ratio.omlx",
   "value": 0.53,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p50 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "strim/omlx",
    "cell": "0.53x"
   }
  },
  {
   "id": "final.p50_agc8.verdict.omlx",
   "value": "NC",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Not claimable: oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog). The summary.md verdict is withheld here because the cell is not claim_eligible.",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p50 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc8.strim",
   "value": 22.67,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "strim",
    "cell": "22.67"
   }
  },
  {
   "id": "final.p95_agc8.rapid-mlx",
   "value": 30.98,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "rapid-mlx",
    "cell": "30.98"
   }
  },
  {
   "id": "final.p95_agc8.mlx-vlm",
   "value": 23.1,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "mlx-vlm",
    "cell": "23.1"
   }
  },
  {
   "id": "final.p95_agc8.mlx-lm",
   "value": 31.68,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "mlx-lm",
    "cell": "31.68"
   }
  },
  {
   "id": "final.p95_agc8.omlx",
   "value": 25.69,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 agent prompts p95 TTFT, omlx. Median of 1 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "1",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "omlx",
    "cell": "25.69"
   }
  },
  {
   "id": "final.p95_agc8.ratio.rapid-mlx",
   "value": 0.73,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.73x"
   }
  },
  {
   "id": "final.p95_agc8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc8.ratio.mlx-vlm",
   "value": 0.98,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.98x"
   }
  },
  {
   "id": "final.p95_agc8.verdict.mlx-vlm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.p95_agc8.ratio.mlx-lm",
   "value": 0.72,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.72x"
   }
  },
  {
   "id": "final.p95_agc8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_agc8.ratio.omlx",
   "value": 0.88,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 agent prompts p95 TTFT, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "strim/omlx",
    "cell": "0.88x"
   }
  },
  {
   "id": "final.p95_agc8.verdict.omlx",
   "value": "NC",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 agent prompts p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Not claimable: oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog). The summary.md verdict is withheld here because the cell is not claim_eligible.",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 agent prompts p95 TTFT",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.a4kc4.strim",
   "value": 13.24,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts aggregate (prefill-bound), strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "strim",
    "cell": "13.24"
   }
  },
  {
   "id": "final.a4kc4.rapid-mlx",
   "value": 13.16,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts aggregate (prefill-bound), rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "rapid-mlx",
    "cell": "13.16"
   }
  },
  {
   "id": "final.a4kc4.mlx-vlm",
   "value": 13.52,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts aggregate (prefill-bound), mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "mlx-vlm",
    "cell": "13.52"
   }
  },
  {
   "id": "final.a4kc4.mlx-lm",
   "value": 13.1,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts aggregate (prefill-bound), mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "mlx-lm",
    "cell": "13.1"
   }
  },
  {
   "id": "final.a4kc4.omlx",
   "value": null,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts aggregate (prefill-bound), omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.a4kc4.ratio.rapid-mlx",
   "value": 1.01,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts aggregate (prefill-bound), Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "strim/rapid-mlx",
    "cell": "1.01x"
   }
  },
  {
   "id": "final.a4kc4.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts aggregate (prefill-bound) against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.a4kc4.ratio.mlx-vlm",
   "value": 0.98,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts aggregate (prefill-bound), Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "strim/mlx-vlm",
    "cell": "0.98x"
   }
  },
  {
   "id": "final.a4kc4.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts aggregate (prefill-bound) against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.a4kc4.ratio.mlx-lm",
   "value": 1.01,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts aggregate (prefill-bound), Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "strim/mlx-lm",
    "cell": "1.01x"
   }
  },
  {
   "id": "final.a4kc4.verdict.mlx-lm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts aggregate (prefill-bound) against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "vs mlx-lm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.a4kc4.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts aggregate (prefill-bound) against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts aggregate (prefill-bound)",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p50_4kc4.strim",
   "value": 10.36,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "strim",
    "cell": "10.36"
   }
  },
  {
   "id": "final.p50_4kc4.rapid-mlx",
   "value": 16.86,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "rapid-mlx",
    "cell": "16.86"
   }
  },
  {
   "id": "final.p50_4kc4.mlx-vlm",
   "value": 16.17,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "mlx-vlm",
    "cell": "16.17"
   }
  },
  {
   "id": "final.p50_4kc4.mlx-lm",
   "value": 16.84,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "mlx-lm",
    "cell": "16.84"
   }
  },
  {
   "id": "final.p50_4kc4.omlx",
   "value": null,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p50 TTFT, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p50_4kc4.ratio.rapid-mlx",
   "value": 0.61,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.61x"
   }
  },
  {
   "id": "final.p50_4kc4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_4kc4.ratio.mlx-vlm",
   "value": 0.64,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.64x"
   }
  },
  {
   "id": "final.p50_4kc4.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_4kc4.ratio.mlx-lm",
   "value": 0.62,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.62x"
   }
  },
  {
   "id": "final.p50_4kc4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_4kc4.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p50 TTFT",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p95_4kc4.strim",
   "value": 16.0,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 0,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "strim",
    "cell": "16"
   }
  },
  {
   "id": "final.p95_4kc4.rapid-mlx",
   "value": 16.86,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "rapid-mlx",
    "cell": "16.86"
   }
  },
  {
   "id": "final.p95_4kc4.mlx-vlm",
   "value": 16.17,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "mlx-vlm",
    "cell": "16.17"
   }
  },
  {
   "id": "final.p95_4kc4.mlx-lm",
   "value": 16.85,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "mlx-lm",
    "cell": "16.85"
   }
  },
  {
   "id": "final.p95_4kc4.omlx",
   "value": null,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=4 4K prompts p95 TTFT, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p95_4kc4.ratio.rapid-mlx",
   "value": 0.95,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.95x"
   }
  },
  {
   "id": "final.p95_4kc4.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_4kc4.ratio.mlx-vlm",
   "value": 0.99,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.99x"
   }
  },
  {
   "id": "final.p95_4kc4.verdict.mlx-vlm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.p95_4kc4.ratio.mlx-lm",
   "value": 0.95,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=4 4K prompts p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.95x"
   }
  },
  {
   "id": "final.p95_4kc4.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_4kc4.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=4 4K prompts p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=4 4K prompts p95 TTFT",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.a4kc8.strim",
   "value": 13.29,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts aggregate (prefill-bound), strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "strim",
    "cell": "13.29"
   }
  },
  {
   "id": "final.a4kc8.rapid-mlx",
   "value": 13.2,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts aggregate (prefill-bound), rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "rapid-mlx",
    "cell": "13.2"
   }
  },
  {
   "id": "final.a4kc8.mlx-vlm",
   "value": 13.84,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts aggregate (prefill-bound), mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "mlx-vlm",
    "cell": "13.84"
   }
  },
  {
   "id": "final.a4kc8.mlx-lm",
   "value": 13.04,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts aggregate (prefill-bound), mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "mlx-lm",
    "cell": "13.04"
   }
  },
  {
   "id": "final.a4kc8.omlx",
   "value": null,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts aggregate (prefill-bound), omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.a4kc8.ratio.rapid-mlx",
   "value": 1.01,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts aggregate (prefill-bound), Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "strim/rapid-mlx",
    "cell": "1.01x"
   }
  },
  {
   "id": "final.a4kc8.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts aggregate (prefill-bound) against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.a4kc8.ratio.mlx-vlm",
   "value": 0.96,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts aggregate (prefill-bound), Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "strim/mlx-vlm",
    "cell": "0.96x"
   }
  },
  {
   "id": "final.a4kc8.verdict.mlx-vlm",
   "value": "LOSS",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts aggregate (prefill-bound) against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "vs mlx-vlm",
    "cell": "LOSS"
   }
  },
  {
   "id": "final.a4kc8.ratio.mlx-lm",
   "value": 1.02,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts aggregate (prefill-bound), Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "strim/mlx-lm",
    "cell": "1.02x"
   }
  },
  {
   "id": "final.a4kc8.verdict.mlx-lm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts aggregate (prefill-bound) against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "vs mlx-lm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.a4kc8.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts aggregate (prefill-bound) against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts aggregate (prefill-bound)",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p50_4kc8.strim",
   "value": 18.75,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "strim",
    "cell": "18.75"
   }
  },
  {
   "id": "final.p50_4kc8.rapid-mlx",
   "value": 26.56,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "rapid-mlx",
    "cell": "26.56"
   }
  },
  {
   "id": "final.p50_4kc8.mlx-vlm",
   "value": 32.07,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "mlx-vlm",
    "cell": "32.07"
   }
  },
  {
   "id": "final.p50_4kc8.mlx-lm",
   "value": 26.83,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "mlx-lm",
    "cell": "26.83"
   }
  },
  {
   "id": "final.p50_4kc8.omlx",
   "value": null,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p50 TTFT, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p50_4kc8.ratio.rapid-mlx",
   "value": 0.71,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.71x"
   }
  },
  {
   "id": "final.p50_4kc8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_4kc8.ratio.mlx-vlm",
   "value": 0.58,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.58x"
   }
  },
  {
   "id": "final.p50_4kc8.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_4kc8.ratio.mlx-lm",
   "value": 0.7,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.70x"
   }
  },
  {
   "id": "final.p50_4kc8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_4kc8.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p50 TTFT",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p95_4kc8.strim",
   "value": 32.0,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 0,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "strim",
    "cell": "32"
   }
  },
  {
   "id": "final.p95_4kc8.rapid-mlx",
   "value": 36.24,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "rapid-mlx",
    "cell": "36.24"
   }
  },
  {
   "id": "final.p95_4kc8.mlx-vlm",
   "value": 32.07,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "mlx-vlm",
    "cell": "32.07"
   }
  },
  {
   "id": "final.p95_4kc8.mlx-lm",
   "value": 36.57,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "mlx-lm",
    "cell": "36.57"
   }
  },
  {
   "id": "final.p95_4kc8.omlx",
   "value": null,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): c=8 4K prompts p95 TTFT, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p95_4kc8.ratio.rapid-mlx",
   "value": 0.88,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.88x"
   }
  },
  {
   "id": "final.p95_4kc8.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_4kc8.ratio.mlx-vlm",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.p95_4kc8.verdict.mlx-vlm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.p95_4kc8.ratio.mlx-lm",
   "value": 0.88,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: c=8 4K prompts p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.88x"
   }
  },
  {
   "id": "final.p95_4kc8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_4kc8.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for c=8 4K prompts p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "c=8 4K prompts p95 TTFT",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.aw5.strim",
   "value": 9.773,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 4 clients, 16K shared prefix aggregate, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "strim",
    "cell": "9.773"
   }
  },
  {
   "id": "final.aw5.rapid-mlx",
   "value": 3.344,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 4 clients, 16K shared prefix aggregate, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "rapid-mlx",
    "cell": "3.344"
   }
  },
  {
   "id": "final.aw5.mlx-vlm",
   "value": 3.473,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 4 clients, 16K shared prefix aggregate, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "mlx-vlm",
    "cell": "3.473"
   }
  },
  {
   "id": "final.aw5.mlx-lm",
   "value": 3.346,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 4 clients, 16K shared prefix aggregate, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 3,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "mlx-lm",
    "cell": "3.346"
   }
  },
  {
   "id": "final.aw5.omlx",
   "value": null,
   "unit": "tok/s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 4 clients, 16K shared prefix aggregate, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.aw5.ratio.rapid-mlx",
   "value": 2.92,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 4 clients, 16K shared prefix aggregate, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "strim/rapid-mlx",
    "cell": "2.92x"
   }
  },
  {
   "id": "final.aw5.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 4 clients, 16K shared prefix aggregate against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aw5.ratio.mlx-vlm",
   "value": 2.81,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 4 clients, 16K shared prefix aggregate, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "strim/mlx-vlm",
    "cell": "2.81x"
   }
  },
  {
   "id": "final.aw5.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 4 clients, 16K shared prefix aggregate against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aw5.ratio.mlx-lm",
   "value": 2.92,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 4 clients, 16K shared prefix aggregate, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "strim/mlx-lm",
    "cell": "2.92x"
   }
  },
  {
   "id": "final.aw5.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 4 clients, 16K shared prefix aggregate against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.aw5.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 4 clients, 16K shared prefix aggregate against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 4 clients, 16K shared prefix aggregate",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p50_w5.strim",
   "value": 20.29,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p50 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "strim",
    "cell": "20.29"
   }
  },
  {
   "id": "final.p50_w5.rapid-mlx",
   "value": 73.53,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p50 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "rapid-mlx",
    "cell": "73.53"
   }
  },
  {
   "id": "final.p50_w5.mlx-vlm",
   "value": 70.48,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p50 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "mlx-vlm",
    "cell": "70.48"
   }
  },
  {
   "id": "final.p50_w5.mlx-lm",
   "value": 73.4,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p50 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "mlx-lm",
    "cell": "73.4"
   }
  },
  {
   "id": "final.p50_w5.omlx",
   "value": null,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p50 TTFT, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p50_w5.ratio.rapid-mlx",
   "value": 0.28,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 p50 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.28x"
   }
  },
  {
   "id": "final.p50_w5.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p50 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_w5.ratio.mlx-vlm",
   "value": 0.29,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 p50 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.29x"
   }
  },
  {
   "id": "final.p50_w5.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p50 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_w5.ratio.mlx-lm",
   "value": 0.28,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 p50 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.28x"
   }
  },
  {
   "id": "final.p50_w5.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p50 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p50_w5.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p50 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p50 TTFT",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p95_w5.strim",
   "value": 22.83,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p95 TTFT, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "strim",
    "cell": "22.83"
   }
  },
  {
   "id": "final.p95_w5.rapid-mlx",
   "value": 73.53,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p95 TTFT, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "rapid-mlx",
    "cell": "73.53"
   }
  },
  {
   "id": "final.p95_w5.mlx-vlm",
   "value": 70.48,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p95 TTFT, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "mlx-vlm",
    "cell": "70.48"
   }
  },
  {
   "id": "final.p95_w5.mlx-lm",
   "value": 73.4,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p95 TTFT, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "mlx-lm",
    "cell": "73.4"
   }
  },
  {
   "id": "final.p95_w5.omlx",
   "value": null,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): W5 p95 TTFT, omlx. No valid value: every pass of this cell was excluded or the engine aborted (n/a in summary.md).",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "n": "0",
   "na": true,
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.p95_w5.ratio.rapid-mlx",
   "value": 0.31,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 p95 TTFT, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "strim/rapid-mlx",
    "cell": "0.31x"
   }
  },
  {
   "id": "final.p95_w5.verdict.rapid-mlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p95 TTFT against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "vs rapid-mlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_w5.ratio.mlx-vlm",
   "value": 0.32,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 p95 TTFT, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "strim/mlx-vlm",
    "cell": "0.32x"
   }
  },
  {
   "id": "final.p95_w5.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p95 TTFT against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_w5.ratio.mlx-lm",
   "value": 0.31,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: W5 p95 TTFT, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "strim/mlx-lm",
    "cell": "0.31x"
   }
  },
  {
   "id": "final.p95_w5.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p95 TTFT against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.p95_w5.verdict.omlx",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for W5 p95 TTFT against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "W5 p95 TTFT",
    "col": "vs omlx",
    "cell": "n/a"
   }
  },
  {
   "id": "final.rp8.strim",
   "value": 48.93,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T8, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "strim",
    "cell": "48.93"
   }
  },
  {
   "id": "final.rp8.rapid-mlx",
   "value": 49.13,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T8, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "rapid-mlx",
    "cell": "49.13"
   }
  },
  {
   "id": "final.rp8.mlx-vlm",
   "value": 192.5,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T8, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "mlx-vlm",
    "cell": "192.5"
   }
  },
  {
   "id": "final.rp8.mlx-lm",
   "value": 55.02,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T8, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "mlx-lm",
    "cell": "55.02"
   }
  },
  {
   "id": "final.rp8.omlx",
   "value": 64.84,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T8, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "omlx",
    "cell": "64.84"
   }
  },
  {
   "id": "final.rp8.ratio.rapid-mlx",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T8, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "strim/rapid-mlx",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.rp8.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T8 against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.rp8.ratio.mlx-vlm",
   "value": 0.25,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T8, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "strim/mlx-vlm",
    "cell": "0.25x"
   }
  },
  {
   "id": "final.rp8.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T8 against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.rp8.ratio.mlx-lm",
   "value": 0.89,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T8, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "strim/mlx-lm",
    "cell": "0.89x"
   }
  },
  {
   "id": "final.rp8.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T8 against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.rp8.ratio.omlx",
   "value": 0.75,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T8, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "strim/omlx",
    "cell": "0.75x"
   }
  },
  {
   "id": "final.rp8.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T8 against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T8",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.rp6.strim",
   "value": 34.6,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T6, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "strim",
    "cell": "34.6"
   }
  },
  {
   "id": "final.rp6.rapid-mlx",
   "value": 34.77,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T6, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "rapid-mlx",
    "cell": "34.77"
   }
  },
  {
   "id": "final.rp6.mlx-vlm",
   "value": 107.7,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T6, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "mlx-vlm",
    "cell": "107.7"
   }
  },
  {
   "id": "final.rp6.mlx-lm",
   "value": 39.31,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T6, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "mlx-lm",
    "cell": "39.31"
   }
  },
  {
   "id": "final.rp6.omlx",
   "value": 44.9,
   "unit": "s",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay prefill sum T1-T6, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 1,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "omlx",
    "cell": "44.9"
   }
  },
  {
   "id": "final.rp6.ratio.rapid-mlx",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T6, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "strim/rapid-mlx",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.rp6.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T6 against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.rp6.ratio.mlx-vlm",
   "value": 0.32,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T6, Strim median divided by mlx-vlm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "strim/mlx-vlm",
    "cell": "0.32x"
   }
  },
  {
   "id": "final.rp6.verdict.mlx-vlm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T6 against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "vs mlx-vlm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.rp6.ratio.mlx-lm",
   "value": 0.88,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T6, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "strim/mlx-lm",
    "cell": "0.88x"
   }
  },
  {
   "id": "final.rp6.verdict.mlx-lm",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T6 against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "vs mlx-lm",
    "cell": "WIN"
   }
  },
  {
   "id": "final.rp6.ratio.omlx",
   "value": 0.77,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay prefill sum T1-T6, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "strim/omlx",
    "cell": "0.77x"
   }
  },
  {
   "id": "final.rp6.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay prefill sum T1-T6 against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay prefill sum T1-T6",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.rpc8.strim",
   "value": 87.25,
   "unit": "%",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay T8 cached fraction, strim. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "Strim build ee9f9d8 (serve defaults: MTP head, admission, batched decode, prefix cache); later commits were not benchmarked.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "strim",
    "cell": "0.8725"
   }
  },
  {
   "id": "final.rpc8.rapid-mlx",
   "value": 87.23,
   "unit": "%",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay T8 cached fraction, rapid-mlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "rapid-mlx 0.15.7, flags: --no-telemetry serve models/qwen38-4bit --port 8651 --max-num-seqs 4 --speculative-config {mtp,qwen38-mtp-4bit}; MTP: yes (--speculative-config mtp, qwen38-mtp-4bit); prefix cache: yes, default on; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "rapid-mlx",
    "cell": "0.8723"
   }
  },
  {
   "id": "final.rpc8.mlx-vlm",
   "value": 0.0,
   "unit": "%",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay T8 cached fraction, mlx-vlm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-vlm 0.7.10, flags: --model models/qwen38-4bit --port 8651 (server defaults); MTP: no; prefix cache: no cross-request reuse (0% cached in replay); rival tuning was not explored.",
   "dataset": "final",
   "dp": 0,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "mlx-vlm",
    "cell": "0"
   }
  },
  {
   "id": "final.rpc8.mlx-lm",
   "value": 87.26,
   "unit": "%",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay T8 cached fraction, mlx-lm. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "mlx-lm 0.32.2, flags: --model models/qwen38-4bit --port 8654 --decode-concurrency 4 --prompt-concurrency 4; MTP: no (no speculative decoding); prefix cache: yes, default in-memory LRU prompt cache; rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "mlx-lm",
    "cell": "0.8726"
   }
  },
  {
   "id": "final.rpc8.omlx",
   "value": 85.25,
   "unit": "%",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Head-to-head 2026-10-28 (h2h-final3-20261028): agent replay T8 cached fraction, omlx. Median of 2 passes (1 rep each); HTTP, one engine at a time, Qwen3.8-27B 4-bit, thinking off, greedy.",
   "caveat": "oMLX 0.4.5.dev1, flags: serve --model-dir <scratch>/models (symlink to models/qwen38-4bit) --port 8653, HOME=<scratch>/home, defaults otherwise (memory-guard balanced, max-concurrent-requests 8); MTP: no (MTP needs oMLX's own checkpoint; stock weights have no MTP head); prefix cache: yes, default paged SSD prefix cache under the scratch HOME (fresh per session, so no cross-session hits); rival tuning was not explored.",
   "dataset": "final",
   "dp": 2,
   "n": "2",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "omlx",
    "cell": "0.8525"
   }
  },
  {
   "id": "final.rpc8.ratio.rapid-mlx",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay T8 cached fraction, Strim median divided by rapid-mlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "strim/rapid-mlx",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.rpc8.verdict.rapid-mlx",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay T8 cached fraction against rapid-mlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "vs rapid-mlx",
    "cell": "TIE"
   }
  },
  {
   "id": "final.rpc8.verdict.mlx-vlm",
   "value": "n/a",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay T8 cached fraction against mlx-vlm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "vs mlx-vlm",
    "cell": "n/a"
   }
  },
  {
   "id": "final.rpc8.ratio.mlx-lm",
   "value": 1.0,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay T8 cached fraction, Strim median divided by mlx-lm median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "strim/mlx-lm",
    "cell": "1.00x"
   }
  },
  {
   "id": "final.rpc8.verdict.mlx-lm",
   "value": "TIE",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay T8 cached fraction against mlx-lm. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "vs mlx-lm",
    "cell": "TIE"
   }
  },
  {
   "id": "final.rpc8.ratio.omlx",
   "value": 1.02,
   "unit": "x",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: agent replay T8 cached fraction, Strim median divided by omlx median (displayed in summary.md).",
   "caveat": "Ratio of two medians of two passes; the TIE band is 2%.",
   "dataset": "final",
   "dp": 2,
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "strim/omlx",
    "cell": "1.02x"
   }
  },
  {
   "id": "final.rpc8.verdict.omlx",
   "value": "WIN",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Final head-to-head: verdict for agent replay T8 cached fraction against omlx. Higher-is-better cells win with ratio above 1, lower-is-better cells (TTFT, time) with ratio below 1; TIE is within 2%. NC = not claimable (claim_eligible is not true for Strim or the rival).",
   "caveat": "Verdicts at the edge of the 2% band can flip with one pass (summary.md lists six such cells).",
   "dataset": "final",
   "verify_md": {
    "section": "Results",
    "row": "agent replay T8 cached fraction",
    "col": "vs omlx",
    "cell": "WIN"
   }
  },
  {
   "id": "final.mem.strim",
   "value": 30.8,
   "unit": "GB",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Peak physical footprint of the engine process tree over the whole run (max over passes and sessions), strim: 30763 / 8711 MB footprint / RSS in summary.md, shown here in GB.",
   "caveat": "phys_footprint includes Metal buffers. Whole-run maximum, not steady state; abort limit was 42 GB.",
   "dataset": "final",
   "dp": 1,
   "verify_md": {
    "section": "Peak memory",
    "row": "**whole run**",
    "col": "strim footprint / RSS",
    "contains": "30763"
   }
  },
  {
   "id": "final.mem.rapid-mlx",
   "value": 35.9,
   "unit": "GB",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Peak physical footprint of the engine process tree over the whole run (max over passes and sessions), rapid-mlx: 35892 / 15640 MB footprint / RSS in summary.md, shown here in GB.",
   "caveat": "phys_footprint includes Metal buffers. Whole-run maximum, not steady state; abort limit was 42 GB.",
   "dataset": "final",
   "dp": 1,
   "verify_md": {
    "section": "Peak memory",
    "row": "**whole run**",
    "col": "rapid-mlx footprint / RSS",
    "contains": "35892"
   }
  },
  {
   "id": "final.mem.mlx-vlm",
   "value": 36.3,
   "unit": "GB",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Peak physical footprint of the engine process tree over the whole run (max over passes and sessions), mlx-vlm: 36313 / 16287 MB footprint / RSS in summary.md, shown here in GB.",
   "caveat": "phys_footprint includes Metal buffers. Whole-run maximum, not steady state; abort limit was 42 GB.",
   "dataset": "final",
   "dp": 1,
   "verify_md": {
    "section": "Peak memory",
    "row": "**whole run**",
    "col": "mlx-vlm footprint / RSS",
    "contains": "36313"
   }
  },
  {
   "id": "final.mem.mlx-lm",
   "value": 37.5,
   "unit": "GB",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Peak physical footprint of the engine process tree over the whole run (max over passes and sessions), mlx-lm: 37450 / 15319 MB footprint / RSS in summary.md, shown here in GB.",
   "caveat": "phys_footprint includes Metal buffers. Whole-run maximum, not steady state; abort limit was 42 GB.",
   "dataset": "final",
   "dp": 1,
   "verify_md": {
    "section": "Peak memory",
    "row": "**whole run**",
    "col": "mlx-lm footprint / RSS",
    "contains": "37450"
   }
  },
  {
   "id": "final.mem.omlx",
   "value": 55.2,
   "unit": "GB",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Peak physical footprint of the engine process tree over the whole run (max over passes and sessions), omlx: 55210 / 18988 MB footprint / RSS in summary.md, shown here in GB.",
   "caveat": "phys_footprint includes Metal buffers. Whole-run maximum, not steady state; abort limit was 42 GB.",
   "dataset": "final",
   "dp": 1,
   "verify_md": {
    "section": "Peak memory",
    "row": "**whole run**",
    "col": "omlx footprint / RSS",
    "contains": "55210"
   }
  },
  {
   "id": "final.agree.self",
   "value": "8/8",
   "unit": "prompts identical",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim pass 1 against Strim pass 2, 8 fixed prompts, 64 greedy tokens.",
   "caveat": "Sanity check only: near-ties diverge, one divergence flips every later token.",
   "dataset": "final",
   "verify": "8/8 identical",
   "verify_section": "Greedy agreement"
  },
  {
   "id": "final.agree.rapid-mlx",
   "value": "7/8",
   "unit": "prompts identical",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim against rapid-mlx, 8 fixed prompts, 64 greedy tokens.",
   "caveat": "Sanity check only: near-ties diverge, one divergence flips every later token.",
   "dataset": "final",
   "verify": "7/8 identical",
   "verify_section": "Greedy agreement"
  },
  {
   "id": "final.agree.mlx-vlm",
   "value": "6/8",
   "unit": "prompts identical",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim against mlx-vlm, 8 fixed prompts, 64 greedy tokens.",
   "caveat": "Sanity check only: near-ties diverge, one divergence flips every later token.",
   "dataset": "final",
   "verify": "6/8 identical",
   "verify_section": "Greedy agreement"
  },
  {
   "id": "final.agree.mlx-lm",
   "value": "6/8",
   "unit": "prompts identical",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim against mlx-lm, 8 fixed prompts, 64 greedy tokens.",
   "caveat": "Sanity check only: near-ties diverge, one divergence flips every later token.",
   "dataset": "final",
   "verify": "6/8 identical",
   "verify_section": "Greedy agreement"
  },
  {
   "id": "final.agree.omlx",
   "value": "7/8",
   "unit": "prompts identical",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim against omlx, 8 fixed prompts, 64 greedy tokens.",
   "caveat": "Sanity check only: near-ties diverge, one divergence flips every later token.",
   "dataset": "final",
   "verify": "7/8 identical",
   "verify_section": "Greedy agreement"
  },
  {
   "id": "final.claim.decode_vs.rapid-mlx",
   "value": "24-50%",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Decode speed over rapid-mlx at 4K, 16K and 32K context: range of strim/rapid-mlx - 1, as a percentage, over the three decode cells (claimable cells only).",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.decode_vs.mlx-vlm",
   "value": "1.8-2.3x",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Decode speed over mlx-vlm at 4K, 16K and 32K context: range of strim/mlx-vlm over the three decode cells (claimable cells only).",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.decode_vs.mlx-lm",
   "value": "1.9-2.5x",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Decode speed over mlx-lm at 4K, 16K and 32K context: range of strim/mlx-lm over the three decode cells (claimable cells only).",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.decode_vs.omlx",
   "value": "1.7-2.2x",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Decode speed over oMLX at 4K, 16K and 32K context: range of strim/omlx over the three decode cells (claimable cells only).",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.w5_speedup",
   "value": "2.8-2.9x",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Aggregate throughput of 4 clients sharing a 16K prefix: range of strim/rival over all rivals.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.ttft_p50_lower",
   "value": "29-56%",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Median first-token latency at 4 and 8 concurrent requests (512-token, agent and 4K prompts): range of 1 - strim/rival over six cells and all rivals.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.agg_lower",
   "value": "2-4%",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Aggregate throughput gap to the best rival at c=4 and c=8 over the six concurrency cells, counting only cells where Strim is more than 2% behind.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.agg_lower_n",
   "value": 3,
   "unit": "cells",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Number of the six aggregate-throughput cells where Strim is more than 2% behind the best rival.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.p95_worse",
   "value": "6%",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "p95 first-token gap to the best rival over the six c=4/c=8 cells, counting only cells where Strim is more than 2% slower.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.p95_worse_n",
   "value": 1,
   "unit": "cells",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Number of the six p95 cells where Strim is more than 2% slower than the best rival.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.prefill_gap",
   "value": "2-4%",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Cold prefill (tok/s) gap to the best rival over 4K, 16K and 32K, counting only sizes where Strim is more than 2% behind.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.prefill_loss",
   "value": "4K, 16K and 32K",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Cold prompt sizes where Strim prefill is more than 2% behind the best rival.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.prefill_tie",
   "value": "none",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Cold prompt sizes where Strim prefill is within 2% of, or ahead of, the best rival.",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.passes",
   "value": 2,
   "unit": "passes",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Passes per engine (pass 1 in the order of the engines list, pass 2 reversed).",
   "caveat": "Recomputed from the displayed ratios by strim site build; there is no hand-typed claim wording.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.claim.build",
   "value": "ee9f9d8",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Strim commit that was benchmarked (from the summary title).",
   "caveat": "Commits after this one were not benchmarked.",
   "dataset": "final",
   "verify": "ee9f9d8",
   "verify_section": null
  },
  {
   "id": "final.val.sessions",
   "value": 30,
   "unit": "sessions",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Engine sessions (engine x pass x cell group) listed in the per-session load table.",
   "caveat": "",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.val.polluted",
   "value": 2,
   "unit": "sessions",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Sessions the summary marks POLLUTED (foreign CPU load above the protocol limit of 3).",
   "caveat": "A polluted session is not a clean measurement; rerun before treating the numbers as final.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.val.clean",
   "value": 28,
   "unit": "sessions",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Primary engine sessions marked CLEAN; later retries are retained separately.",
   "caveat": "Claims additionally require claim_eligible for both engines in each cell.",
   "dataset": "final",
   "computed": true
  },
  {
   "id": "final.val.load_max",
   "value": 4.68,
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Highest 1-minute load average recorded at a cell start, over all sessions (per-session table).",
   "caveat": "Includes the engine's own load; the protocol limit for a clean session is 3 from non-engine work.",
   "dataset": "final",
   "dp": 1,
   "computed": true
  },
  {
   "id": "final.claim.swap",
   "value": 0,
   "unit": "MB",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Largest swap growth in any benchmark session.",
   "caveat": "Baseline swap was pre-existing.",
   "dataset": "final",
   "verify": "swap growth max 0 MB",
   "verify_section": "Run conditions"
  },
  {
   "id": "ver.h2h.rapid-mlx",
   "value": "rapid-mlx 0.15.7",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Measured version of rapid-mlx in the head-to-head run (engine_meta.version in results.json, listed in the summary).",
   "caveat": "",
   "dataset": "final",
   "verify": "0.15.7"
  },
  {
   "id": "ver.h2h.mlx-vlm",
   "value": "mlx-vlm 0.7.10",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Measured version of mlx-vlm in the head-to-head run (engine_meta.version in results.json, listed in the summary).",
   "caveat": "",
   "dataset": "final",
   "verify": "0.7.10"
  },
  {
   "id": "ver.h2h.mlx-lm",
   "value": "mlx-lm 0.32.2",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Measured version of mlx-lm in the head-to-head run (engine_meta.version in results.json, listed in the summary).",
   "caveat": "",
   "dataset": "final",
   "verify": "0.32.2"
  },
  {
   "id": "ver.h2h.omlx",
   "value": "oMLX 0.4.5.dev1",
   "unit": "",
   "source_path": "bench/results/h2h-final3-20261028/summary.md",
   "method_note": "Measured version of omlx in the head-to-head run (engine_meta.version in results.json, listed in the summary).",
   "caveat": "",
   "dataset": "final",
   "verify": "0.4.5.dev1"
  }
 ],
 "final": {
  "source": "bench/results/h2h-final3-20261028/summary.md",
  "engines": [
   "strim",
   "rapid-mlx",
   "mlx-vlm",
   "mlx-lm",
   "omlx"
  ],
  "not_claimable": {
   "aagc4": {
    "omlx": "oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p50_agc4": {
    "omlx": "oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p95_agc4": {
    "omlx": "oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "aagc8": {
    "omlx": "oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p50_agc8": {
    "omlx": "oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p95_agc8": {
    "omlx": "oMLX: pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "a4kc4": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p50_4kc4": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p95_4kc4": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "a4kc8": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p50_4kc8": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p95_4kc8": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "aw5": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p50_w5": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   },
   "p95_w5": {
    "omlx": "oMLX: no valid value, pass 1 excluded (engine_guard_abort, memory_abort_watchdog), pass 2 excluded (engine_guard_abort, memory_abort_watchdog)"
   }
  },
  "cells": [
   {
    "id": "c4k_pf",
    "label": "cold 4K prefill",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 994.516,
     "rapid-mlx": 996.793,
     "mlx-vlm": 1016.76,
     "mlx-lm": 949.083,
     "omlx": 896.71
    },
    "per_pass": {
     "strim": [
      990.2591763647282,
      998.7737541188864
     ],
     "rapid-mlx": [
      993.5955083481768,
      999.9908488769455
     ],
     "mlx-vlm": [
      1015.6156469704706,
      1017.8993544220102
     ],
     "mlx-lm": [
      948.9737704784246,
      949.1914573317958
     ],
     "omlx": [
      896.2131638654869,
      897.2074989399104
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "c4k_tt",
    "label": "cold 4K TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 4.12972,
     "rapid-mlx": 4.12026,
     "mlx-vlm": 4.03932,
     "mlx-lm": 4.32734,
     "omlx": 4.58008
    },
    "per_pass": {
     "strim": [
      4.147399083012715,
      4.112042375025339
     ],
     "rapid-mlx": [
      4.133472791989334,
      4.107037584006321
     ],
     "mlx-vlm": [
      4.0438526250072755,
      4.034780042013153
     ],
     "mlx-lm": [
      4.327832999988459,
      4.326840458030347
     ],
     "omlx": [
      4.582615125051234,
      4.577536416996736
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "d4k",
    "label": "decode 256 tok @ 4K ctx",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 75.3386,
     "rapid-mlx": 50.1205,
     "mlx-vlm": 32.6902,
     "mlx-lm": 30.1058,
     "omlx": 33.5263
    },
    "per_pass": {
     "strim": [
      74.64110888676724,
      76.03612735705855
     ],
     "rapid-mlx": [
      48.90937840234158,
      51.33163401688596
     ],
     "mlx-vlm": [
      32.719537096804686,
      32.66081025175724
     ],
     "mlx-lm": [
      30.079984597115736,
      30.1316226093821
     ],
     "omlx": [
      33.533956805267124,
      33.518641581396196
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "c16k_pf",
    "label": "cold 16K prefill",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 963.081,
     "rapid-mlx": 972.366,
     "mlx-vlm": 988.039,
     "mlx-lm": 919.824,
     "omlx": 891.921
    },
    "per_pass": {
     "strim": [
      956.0441104955673,
      970.1182379757528
     ],
     "rapid-mlx": [
      967.9860365197084,
      976.7453096939153
     ],
     "mlx-vlm": [
      987.575208063232,
      988.5027667653405
     ],
     "mlx-lm": [
      919.8360799524447,
      919.8113156098983
     ],
     "omlx": [
      892.1081525642752,
      891.733113908098
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "c16k_tt",
    "label": "cold 16K TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 17.0244,
     "rapid-mlx": 16.8613,
     "mlx-vlm": 16.5935,
     "mlx-lm": 17.8241,
     "omlx": 18.3817
    },
    "per_pass": {
     "strim": [
      17.148790333012585,
      16.900001832982525
     ],
     "rapid-mlx": [
      16.937227791990153,
      16.78533783298917
     ],
     "mlx-vlm": [
      16.601267291989643,
      16.58568954101065
     ],
     "mlx-lm": [
      17.823827915999573,
      17.82430779200513
     ],
     "omlx": [
      18.377816582971718,
      18.385545791999903
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "d16k",
    "label": "decode 256 tok @ 16K ctx",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 65.1795,
     "rapid-mlx": 45.749,
     "mlx-vlm": 30.934,
     "mlx-lm": 28.6754,
     "omlx": 31.4219
    },
    "per_pass": {
     "strim": [
      62.46886891454122,
      67.89015071310035
     ],
     "rapid-mlx": [
      45.448993280541536,
      46.04895247400194
     ],
     "mlx-vlm": [
      30.965347284484512,
      30.902559074388684
     ],
     "mlx-lm": [
      28.737987478095178,
      28.612712763106046
     ],
     "omlx": [
      31.407909585455236,
      31.4358064190046
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "c32k_pf",
    "label": "cold 32K prefill",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 884.629,
     "rapid-mlx": 910.005,
     "mlx-vlm": 923.193,
     "mlx-lm": 850.827,
     "omlx": 823.739
    },
    "per_pass": {
     "strim": [
      872.6838022697696,
      896.574278148532
     ],
     "rapid-mlx": [
      906.9449453053962,
      913.0643932415978
     ],
     "mlx-vlm": [
      922.9500087062694,
      923.4365616009732
     ],
     "mlx-lm": [
      850.2966577741058,
      851.3581738196085
     ],
     "omlx": [
      822.5779456564721,
      824.8995013191036
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "c32k_tt",
    "label": "cold 32K TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 37.0607,
     "rapid-mlx": 36.0211,
     "mlx-vlm": 35.5061,
     "mlx-lm": 38.526,
     "omlx": 39.793
    },
    "per_pass": {
     "strim": [
      37.56114175001858,
      36.56027258298127
     ],
     "rapid-mlx": [
      36.14221587503562,
      35.89998716698028
     ],
     "mlx-vlm": [
      35.51546637498541,
      35.496753499959596
     ],
     "mlx-lm": [
      38.55007508298149,
      38.502008917042986
     ],
     "omlx": [
      39.84911116701551,
      39.736961833026726
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "d32k",
    "label": "decode 128 tok @ 32K ctx",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 52.1703,
     "rapid-mlx": 42.169,
     "mlx-vlm": 29.349,
     "mlx-lm": 27.6153,
     "omlx": 29.9042
    },
    "per_pass": {
     "strim": [
      49.109589812569176,
      55.23104196732022
     ],
     "rapid-mlx": [
      42.36726480988293,
      41.970641127312646
     ],
     "mlx-vlm": [
      29.44298765099528,
      29.25505879753046
     ],
     "mlx-lm": [
      27.611249194518408,
      27.61934886731594
     ],
     "omlx": [
      29.932424308882194,
      29.87603357654898
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "a512c4",
    "label": "c=4 512-tok prompts aggregate",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 69.217,
     "rapid-mlx": 71.6904,
     "mlx-vlm": 67.294,
     "mlx-lm": 66.3094,
     "omlx": 54.6131
    },
    "per_pass": {
     "strim": [
      68.89521764334172,
      69.53888088183697
     ],
     "rapid-mlx": [
      71.62508451056928,
      71.75576984087522
     ],
     "mlx-vlm": [
      67.52973790330746,
      67.05832885217677
     ],
     "mlx-lm": [
      66.41515530108697,
      66.2036084923877
     ],
     "omlx": [
      53.922727471503364,
      55.303456567038076
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "p50_512c4",
    "label": "c=4 512-tok prompts p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 1.4931,
     "rapid-mlx": 2.33915,
     "mlx-vlm": 2.30482,
     "mlx-lm": 2.4667,
     "omlx": 3.04904
    },
    "per_pass": {
     "strim": [
      1.4997049160301685,
      1.4864987709734123
     ],
     "rapid-mlx": [
      2.3466427079983987,
      2.3316598965320736
     ],
     "mlx-vlm": [
      2.299901812482858,
      2.3097471255168784
     ],
     "mlx-lm": [
      2.4666102085029706,
      2.46678129202337
     ],
     "omlx": [
      3.173189374996582,
      2.9248848960269243
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "p95_512c4",
    "label": "c=4 512-tok prompts p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 2.29205,
     "rapid-mlx": 2.33916,
     "mlx-vlm": 2.30494,
     "mlx-lm": 2.46895,
     "omlx": 3.19772
    },
    "per_pass": {
     "strim": [
      2.2991381722327784,
      2.2849593661318064
     ],
     "rapid-mlx": [
      2.3466546223295155,
      2.331661468811217
     ],
     "mlx-vlm": [
      2.300043581263162,
      2.3098422063689212
     ],
     "mlx-lm": [
      2.4690242563869105,
      2.4688733517949
     ],
     "omlx": [
      3.324441466142889,
      3.070992528620991
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "a512c8",
    "label": "c=8 512-tok prompts aggregate",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 76.3713,
     "rapid-mlx": 71.7601,
     "mlx-vlm": 74.6549,
     "mlx-lm": 67.8957,
     "omlx": 58.7399
    },
    "per_pass": {
     "strim": [
      76.06283677969233,
      76.67966851303228
     ],
     "rapid-mlx": [
      71.66717184498926,
      71.85296406924105
     ],
     "mlx-vlm": [
      74.88146730994514,
      74.42840650970386
     ],
     "mlx-lm": [
      67.88206292729927,
      67.90929719558979
     ],
     "omlx": [
      58.639098334836454,
      58.8406402821852
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "p50_512c8",
    "label": "c=8 512-tok prompts p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 2.6835,
     "rapid-mlx": 5.88939,
     "mlx-vlm": 4.34047,
     "mlx-lm": 6.0643,
     "omlx": 5.72625
    },
    "per_pass": {
     "strim": [
      2.720974333002232,
      2.646025207999628
     ],
     "rapid-mlx": [
      5.901846958498936,
      5.876939062029123
     ],
     "mlx-vlm": [
      4.322359583486104,
      4.358571604505414
     ],
     "mlx-lm": [
      6.073393354541622,
      6.055213812505826
     ],
     "omlx": [
      5.753753083001357,
      5.69875633349875
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "p95_512c8",
    "label": "c=8 512-tok prompts p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 4.60995,
     "rapid-mlx": 9.4659,
     "mlx-vlm": 4.3407,
     "mlx-lm": 9.83719,
     "omlx": 6.30375
    },
    "per_pass": {
     "strim": [
      4.657134281177423,
      4.562765900004887
     ],
     "rapid-mlx": [
      9.479474070490687,
      9.452318600111175
     ],
     "mlx-vlm": [
      4.322517264279304,
      4.358890007913578
     ],
     "mlx-lm": [
      9.853219283002545,
      9.821168749462231
     ],
     "omlx": [
      6.345935666884179,
      6.261567833437584
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "aagc4",
    "label": "c=4 agent prompts (~2.8K tok) aggregate",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 28.7283,
     "rapid-mlx": 27.729,
     "mlx-vlm": 28.5687,
     "mlx-lm": 27.2156,
     "omlx": 25.277022684698338
    },
    "per_pass": {
     "strim": [
      28.7369845524891,
      28.71951967036254
     ],
     "rapid-mlx": [
      27.70149047006139,
      27.756607754912185
     ],
     "mlx-vlm": [
      28.647259215847495,
      28.49009302515629
     ],
     "mlx-lm": [
      27.147750311128263,
      27.283398050485438
     ],
     "omlx": [
      25.277022684698338
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p50_agc4",
    "label": "c=4 agent prompts p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 7.67754,
     "rapid-mlx": 13.439,
     "mlx-vlm": 12.4301,
     "mlx-lm": 13.4678,
     "omlx": 13.31877581300796
    },
    "per_pass": {
     "strim": [
      7.695195166481426,
      7.6598812914744485
     ],
     "rapid-mlx": [
      13.456243604508927,
      13.42169252052554
     ],
     "mlx-vlm": [
      12.395282979006879,
      12.465013937995536
     ],
     "mlx-lm": [
      13.498417625029106,
      13.437261562503409
     ],
     "omlx": [
      13.31877581300796
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p95_agc4",
    "label": "c=4 agent prompts p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 12.0733,
     "rapid-mlx": 13.439,
     "mlx-vlm": 12.4302,
     "mlx-lm": 13.4696,
     "omlx": 13.472804481285857
    },
    "per_pass": {
     "strim": [
      12.118655014241812,
      12.027920441940658
     ],
     "rapid-mlx": [
      13.456259006261826,
      13.421711412494187
     ],
     "mlx-vlm": [
      12.395381018627086,
      12.465060604567407
     ],
     "mlx-lm": [
      13.500029979436658,
      13.43920824554516
     ],
     "omlx": [
      13.472804481285857
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "aagc8",
    "label": "c=8 agent prompts aggregate",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 31.1928,
     "rapid-mlx": 28.4732,
     "mlx-vlm": 31.1774,
     "mlx-lm": 27.6328,
     "omlx": 26.893094469953493
    },
    "per_pass": {
     "strim": [
      31.058883791666254,
      31.32669517706924
     ],
     "rapid-mlx": [
      28.102641095604362,
      28.84378155887728
     ],
     "mlx-vlm": [
      31.25065339980182,
      31.104077976614192
     ],
     "mlx-lm": [
      27.598599190289164,
      27.66700882963603
     ],
     "omlx": [
      26.893094469953493
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p50_agc8",
    "label": "c=8 agent prompts p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 13.3427,
     "rapid-mlx": 22.0974,
     "mlx-vlm": 23.0982,
     "mlx-lm": 22.6107,
     "omlx": 25.079944395984057
    },
    "per_pass": {
     "strim": [
      13.322275937505765,
      13.363075375498738
     ],
     "rapid-mlx": [
      22.204041874501854,
      21.990672020998318
     ],
     "mlx-vlm": [
      23.06750191701576,
      23.128884103993187
     ],
     "mlx-lm": [
      22.63748277098057,
      22.584005479497137
     ],
     "omlx": [
      25.079944395984057
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p95_agc8",
    "label": "c=8 agent prompts p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 22.6717,
     "rapid-mlx": 30.9825,
     "mlx-vlm": 23.0995,
     "mlx-lm": 31.6794,
     "omlx": 25.685719004581916
    },
    "per_pass": {
     "strim": [
      22.820181774467343,
      22.523155427191522
     ],
     "rapid-mlx": [
      31.455370599971502,
      30.509650635209983
     ],
     "mlx-vlm": [
      23.06985785975703,
      23.129135960663554
     ],
     "mlx-lm": [
      31.71786625845125,
      31.640849556209286
     ],
     "omlx": [
      25.685719004581916
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "a4kc4",
    "label": "c=4 4K prompts aggregate (prefill-bound)",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 13.2369,
     "rapid-mlx": 13.1613,
     "mlx-vlm": 13.5173,
     "mlx-lm": 13.0998
    },
    "per_pass": {
     "strim": [
      13.28056585263768,
      13.193249397770483
     ],
     "rapid-mlx": [
      13.096246090228522,
      13.2262970561705
     ],
     "mlx-vlm": [
      13.553951633802114,
      13.480556781413014
     ],
     "mlx-lm": [
      13.083382242953997,
      13.116264563248466
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p50_4kc4",
    "label": "c=4 4K prompts p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 10.3625,
     "rapid-mlx": 16.8558,
     "mlx-vlm": 16.1717,
     "mlx-lm": 16.8446
    },
    "per_pass": {
     "strim": [
      10.389784458006034,
      10.335266291513108
     ],
     "rapid-mlx": [
      16.89036414551083,
      16.821316666493658
     ],
     "mlx-vlm": [
      16.129387896013213,
      16.21398506249534
     ],
     "mlx-lm": [
      16.865549792011734,
      16.82363620851538
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p95_4kc4",
    "label": "c=4 4K prompts p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 15.9992,
     "rapid-mlx": 16.8559,
     "mlx-vlm": 16.1717,
     "mlx-lm": 16.8462
    },
    "per_pass": {
     "strim": [
      16.04394589991134,
      15.954539590200874
     ],
     "rapid-mlx": [
      16.890381424996303,
      16.821354568810783
     ],
     "mlx-vlm": [
      16.12939299554273,
      16.21403056695417
     ],
     "mlx-lm": [
      16.867203677634826,
      16.825282172395966
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "a4kc8",
    "label": "c=8 4K prompts aggregate (prefill-bound)",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 13.2936,
     "rapid-mlx": 13.2033,
     "mlx-vlm": 13.8365,
     "mlx-lm": 13.0407
    },
    "per_pass": {
     "strim": [
      13.209520210758383,
      13.377597446174038
     ],
     "rapid-mlx": [
      13.192249864973292,
      13.214348378334625
     ],
     "mlx-vlm": [
      13.892911435474106,
      13.780100087362115
     ],
     "mlx-lm": [
      13.045700355982651,
      13.035669948704655
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p50_4kc8",
    "label": "c=8 4K prompts p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 18.7503,
     "rapid-mlx": 26.5585,
     "mlx-vlm": 32.0735,
     "mlx-lm": 26.8316
    },
    "per_pass": {
     "strim": [
      18.8558203119901,
      18.64470752052148
     ],
     "rapid-mlx": [
      26.58756104198983,
      26.52938443797757
     ],
     "mlx-vlm": [
      31.94380550002097,
      32.203213520988356
     ],
     "mlx-lm": [
      26.816113750013756,
      26.84702549999929
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p95_4kc8",
    "label": "c=8 4K prompts p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 31.9956,
     "rapid-mlx": 36.2376,
     "mlx-vlm": 32.0748,
     "mlx-lm": 36.5658
    },
    "per_pass": {
     "strim": [
      32.192702170572005,
      31.7984509710659
     ],
     "rapid-mlx": [
      36.26950996856031,
      36.20563456700474
     ],
     "mlx-vlm": [
      31.944214449467836,
      32.20543624583515
     ],
     "mlx-lm": [
      36.54713623693387,
      36.584527047676964
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "aw5",
    "label": "W5 4 clients, 16K shared prefix aggregate",
    "unit": "tok/s",
    "better": "higher",
    "median": {
     "strim": 9.77319,
     "rapid-mlx": 3.34356,
     "mlx-vlm": 3.47338,
     "mlx-lm": 3.34618
    },
    "per_pass": {
     "strim": [
      9.734546028038928,
      9.811841537422566
     ],
     "rapid-mlx": [
      3.3424440502530754,
      3.3446849403395285
     ],
     "mlx-vlm": [
      3.5021039621845165,
      3.444648421597538
     ],
     "mlx-lm": [
      3.3415622312427806,
      3.3507904754004794
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p50_w5",
    "label": "W5 p50 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 20.2925,
     "rapid-mlx": 73.5338,
     "mlx-vlm": 70.4808,
     "mlx-lm": 73.3956
    },
    "per_pass": {
     "strim": [
      20.365314187511103,
      20.219642292009667
     ],
     "rapid-mlx": [
      73.53973824999412,
      73.5279119584884
     ],
     "mlx-vlm": [
      69.88661229147692,
      71.07506408300833
     ],
     "mlx-lm": [
      73.49170729148318,
      73.29942797950935
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "p95_w5",
    "label": "W5 p95 TTFT",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 22.8346,
     "rapid-mlx": 73.5339,
     "mlx-vlm": 70.4809,
     "mlx-lm": 73.3972
    },
    "per_pass": {
     "strim": [
      22.90824494556,
      22.760883662544074
     ],
     "rapid-mlx": [
      73.53982229820394,
      73.52796024824492
     ],
     "mlx-vlm": [
      69.88667804813304,
      71.07512262678647
     ],
     "mlx-lm": [
      73.49342685361334,
      73.30098296257783
     ],
     "omlx": []
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": false
    }
   },
   {
    "id": "rp8",
    "label": "agent replay prefill sum T1-T8",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 48.9337,
     "rapid-mlx": 49.1317,
     "mlx-vlm": 192.494,
     "mlx-lm": 55.0172,
     "omlx": 64.8438
    },
    "per_pass": {
     "strim": [
      49.04054024798097,
      48.82678403996397
     ],
     "rapid-mlx": [
      49.21724704094231,
      49.04606283304747
     ],
     "mlx-vlm": [
      192.11589604103938,
      192.87292145896936
     ],
     "mlx-lm": [
      56.15164162393194,
      53.88273254100932
     ],
     "omlx": [
      64.92123408301268,
      64.76629537489498
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "rp6",
    "label": "agent replay prefill sum T1-T6",
    "unit": "s",
    "better": "lower",
    "median": {
     "strim": 34.5958,
     "rapid-mlx": 34.7669,
     "mlx-vlm": 107.681,
     "mlx-lm": 39.3085,
     "omlx": 44.8998
    },
    "per_pass": {
     "strim": [
      34.67593983199913,
      34.51573699899018
     ],
     "rapid-mlx": [
      34.808584456914105,
      34.72528483305359
     ],
     "mlx-vlm": [
      107.26883254101267,
      108.09303366695531
     ],
     "mlx-lm": [
      39.36328779091127,
      39.25380787398899
     ],
     "omlx": [
      44.90104391699424,
      44.89864512486383
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   },
   {
    "id": "rpc8",
    "label": "agent replay T8 cached fraction",
    "unit": "%",
    "better": "higher",
    "median": {
     "strim": 0.872475,
     "rapid-mlx": 0.872285,
     "mlx-vlm": 0.0,
     "mlx-lm": 0.87263,
     "omlx": 0.852487
    },
    "per_pass": {
     "strim": [
      0.8724748402262543,
      0.8724748402262543
     ],
     "rapid-mlx": [
      0.8722820763956906,
      0.8722883306400274
     ],
     "mlx-vlm": [
      0.0,
      0.0
     ],
     "mlx-lm": [
      0.8726279963761906,
      0.872631115028647
     ],
     "omlx": [
      0.8524975514201762,
      0.8524766778482408
     ]
    },
    "claim_eligible": {
     "strim": true,
     "rapid-mlx": true,
     "mlx-vlm": true,
     "mlx-lm": true,
     "omlx": true
    }
   }
  ],
  "validity": {
   "run": "h2h-final3-20261028",
   "date": "2026-10-28",
   "build": "ee9f9d8",
   "clean": false,
   "sessions": [
    {
     "engine": "mlx-lm",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 2.05,
     "end": 1793223880
    },
    {
     "engine": "mlx-lm",
     "pass": 1,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 1.1,
     "end": 1793224311
    },
    {
     "engine": "mlx-lm",
     "pass": 1,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.51,
     "end": 1793224500
    },
    {
     "engine": "mlx-lm",
     "pass": 2,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 1.48,
     "end": 1793226431
    },
    {
     "engine": "mlx-lm",
     "pass": 2,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 1.5,
     "end": 1793226711
    },
    {
     "engine": "mlx-lm",
     "pass": 2,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.41,
     "end": 1793226899
    },
    {
     "engine": "mlx-vlm",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 2.89,
     "end": 1793223036
    },
    {
     "engine": "mlx-vlm",
     "pass": 1,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "POLLUTED",
     "load_max": 1.84,
     "end": 1793223219
    },
    {
     "engine": "mlx-vlm",
     "pass": 1,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.68,
     "end": 1793223547
    },
    {
     "engine": "mlx-vlm",
     "pass": 1,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 1.92,
     "end": 1793229709
    },
    {
     "engine": "mlx-vlm",
     "pass": 2,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 1.85,
     "end": 1793227226
    },
    {
     "engine": "mlx-vlm",
     "pass": 2,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 3.69,
     "end": 1793227408
    },
    {
     "engine": "mlx-vlm",
     "pass": 2,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 2.72,
     "end": 1793227738
    },
    {
     "engine": "omlx",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 2.76,
     "end": 1793224834
    },
    {
     "engine": "omlx",
     "pass": 1,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 2.78,
     "end": 1793225074
    },
    {
     "engine": "omlx",
     "pass": 1,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.92,
     "end": 1793225299
    },
    {
     "engine": "omlx",
     "pass": 2,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 1.74,
     "end": 1793225633
    },
    {
     "engine": "omlx",
     "pass": 2,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 2.39,
     "end": 1793225873
    },
    {
     "engine": "omlx",
     "pass": 2,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.74,
     "end": 1793226097
    },
    {
     "engine": "rapid-mlx",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "POLLUTED",
     "load_max": 2.94,
     "end": 1793222301
    },
    {
     "engine": "rapid-mlx",
     "pass": 1,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 4.68,
     "end": 1793222520
    },
    {
     "engine": "rapid-mlx",
     "pass": 1,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.9,
     "end": 1793222707
    },
    {
     "engine": "rapid-mlx",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "POLLUTED",
     "load_max": 1.5,
     "end": 1793229320
    },
    {
     "engine": "rapid-mlx",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 1.76,
     "end": 1793235937
    },
    {
     "engine": "rapid-mlx",
     "pass": 2,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 1.69,
     "end": 1793228055
    },
    {
     "engine": "rapid-mlx",
     "pass": 2,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 1.86,
     "end": 1793228243
    },
    {
     "engine": "rapid-mlx",
     "pass": 2,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.83,
     "end": 1793228429
    },
    {
     "engine": "strim",
     "pass": 1,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 2.53,
     "end": 1793221310
    },
    {
     "engine": "strim",
     "pass": 1,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 3.63,
     "end": 1793221671
    },
    {
     "engine": "strim",
     "pass": 1,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 2.44,
     "end": 1793221802
    },
    {
     "engine": "strim",
     "pass": 2,
     "cells": [
      "agree",
      "w3",
      "w3_32k"
     ],
     "status": "CLEAN",
     "load_max": 1.88,
     "end": 1793228736
    },
    {
     "engine": "strim",
     "pass": 2,
     "cells": [
      "w4",
      "w4s",
      "w6"
     ],
     "status": "CLEAN",
     "load_max": 1.52,
     "end": 1793228916
    },
    {
     "engine": "strim",
     "pass": 2,
     "cells": [
      "w5",
      "replay"
     ],
     "status": "CLEAN",
     "load_max": 1.34,
     "end": 1793229046
    }
   ]
  }
 }
}