{
  "schema_version": 1,
  "generated_at_utc": "2026-08-04T07:09:35.288126+00:00",
  "note": "Dashboard headline data for site/app.js's hero results band. Only raw measured tok/s values are stored here -- every ratio shown on the dashboard is computed from these numbers at render time (see site/app.js renderHeadline()), so the page cannot drift from the source docs below without this file being re-derived from them.",
  "source_docs": [
    "results/AUTODEFAULTS.md",
    "results/REMEASURE-2026-08-04-QUIET.md",
    "README.md (section: \"Decomposition — how much of the decode win is SME2, and how much is thread tuning\")"
  ],
  "host": "Apple M4 Max, macOS",
  "model": "Qwen2.5-0.5B-Instruct-Q4_0 (/tmp/ggufs/q05.gguf)",
  "patch": "patches/0002-kleidiai-sme-aware-thread-default.patch",
  "kill_switch_env": "GGML_KLEIDIAI_AUTO_THREADS=0",
  "autodefaults": {
    "source": "results/AUTODEFAULTS.md §5 -- round-robin interleaved (1,2,3,4,1,2,3,4,...), n=9 per config, llama-cli, decode and prefill measured separately from a single invocation per rep",
    "unit": "tok/s",
    "decode": {
      "baseline_no_flags": 67.8,
      "autodefault_no_flags": 145.9,
      "hand_tuned_t2": 146.0,
      "kill_switch_no_flags": 68.3
    },
    "prefill": {
      "baseline_no_flags": 1835.2,
      "autodefault_no_flags": 1779.8,
      "naive_t2_workaround": 975.6,
      "kill_switch_no_flags": 1835.3
    }
  },
  "tuning_study": {
    "source": "results/REMEASURE-2026-08-04-QUIET.md §Results -- round-robin interleaved, n=7 per config, llama-bench, quiet machine (superseding the earlier contended measurement)",
    "unit": "tok/s",
    "decode": { "default_no_flags": 93.6, "tuned_t2": 321.0 },
    "prefill": { "default_no_flags": 1230.3, "tuned_t8": 2198.1 }
  },
  "decomposition": {
    "source": "README.md, §'Decomposition — how much of the decode win is SME2, and how much is thread tuning' -- a second sweep toggling thread count AND GGML_KLEIDIAI_SME together, interleaved n=5, same session. Absolute tok/s here were collected under different shared-machine load than the quiet tuning_study sweep above and must never be compared to it directly -- only the ratios within this one interleaved session are load-bearing (this project's own interleaving-discipline caveat).",
    "unit": "tok/s",
    "default_threads_sme_on": 48.0,
    "default_threads_sme_off": 59.6,
    "t2_sme_on": 309.2,
    "t2_sme_off": 235.7
  }
}