{
  "schema": "s2s-bench-v1",
  "exhibit": "two-hours-on-battery",
  "licence": "CC BY 4.0",
  "attribution": "strata\u2192signal research, research.strata2signal.com",
  "contact": "hello@strata2signal.com",
  "run_id": "cpu-only-0825",
  "run_window_utc": {
    "main_bench_opened_utc": "2026-08-25T14:09:19Z",
    "main_bench_closed_utc": "2026-08-25T17:17:57Z",
    "supplement_first_record_utc": "2026-08-25T17:57:19Z",
    "supplement_closed_utc": "2026-08-25T19:45:39Z",
    "note": "Each arm carries its own started_utc / finished_utc. The supplement's first stamp is a FINISH time - the first sweep arm does not record its own start, so the supplement began somewhat before it. The main window's opening arm is cpu-mini-m1; the arm that started at 14:32:03Z is the one the bench discarded."
  },
  "machines": {
    "cpu-server": "a server with 128G of ECC RAM and a server-class processor: 64 logical CPUs, 32 physical, not a hybrid part; a GPU present in the box and parked",
    "cpu-laptop": "test Rig B, a laptop: 24 logical CPUs - 8 performance cores with no hyperthreading plus 16 efficiency cores; hybrid Intel part; a 24G-VRAM-class GPU present and parked",
    "cpu-mini": "a mini PC with 32G RAM: 16 logical CPUs - 4 performance cores with hyperthreading (8 performance logical CPUs) plus 8 efficiency cores; hybrid Intel part; no discrete GPU at all"
  },
  "counting_rules": {
    "decode": "Median of the five scored decode_tok_s values (three for the 70B), warmup discarded. The warmup stays in the file.",
    "prefill": "COLD runs only. The warm-run prefill field is a prefix-cache artifact reporting 2,613 tok/s where the truth is 33.6, and is never tabulated.",
    "latency": "Four distinct numbers, never one 'TTFT': cached-prefix, fresh-prompt-model-loaded, fully-cold, and decode. They differ by up to 96x, so the multiplier's denominator is always named.",
    "threads": "Ollama's default gave both hybrid boxes 4 threads. Every cross-machine ratio here is configuration-inclusive, not silicon-pure, and the kit says so rather than normalising it away.",
    "spread": "TWO different denominators: dispersion is (max - min) over the MEDIAN; the bandwidth spread is (best - worst) over the WORST. Both are stated where they stand.",
    "gpu_receipt": "Three fields per arm - size_vram 0, cpu_only_receipt true, and the nvidia-smi snapshot compared byte for byte before and after. Twenty of the 32 files in runs/ carry all three; nineteen of those twenty are the scored decode-and-prefill arms and all nineteen have identical snapshots. The twentieth is the discarded arm, whose snapshot changed."
  },
  "not_a_standard": "Six models, three machines, one day, one prompt, one context size, one quantisation. Sized to one question and not a leaderboard, not a benchmark suite, and not a quality ranking of anything.",
  "kind": "data-kit-index",
  "exhibit_title": "Two Hours at 12 tok/s — On Battery",
  "exhibit_url": "https://research.strata2signal.com/two-hours-on-battery/index.html",
  "generated_utc": "2026-08-27T00:00:00Z",
  "sanitisation": "Sanitised at the pen, never copied raw. provenance.json names every rule applied to every file, with the count of times each one fired, and names what was left out and why.",
  "file_count": 74,
  "files_listed": 71,
  "files": [
    {
      "file": "README.md",
      "role": "What each file is, the machines by class, the five counting rules in short form, the timestamp conversion, what is deliberately not here, and how to redo the run yourself.",
      "media_type": "text/markdown",
      "bytes": 14086,
      "sha256": "bc05089de90d295bd6eecda2417c704472a09fed1f90f9f5e04dc1d2a6b46473",
      "rules_applied": [
        "name-the-omission"
      ]
    },
    {
      "file": "NOTES.md",
      "role": "The measurement record: every headline figure on the exhibit page with the command that re-derives it from this kit, including the two the page states as ranges.",
      "media_type": "text/markdown",
      "bytes": 15436,
      "sha256": "577d7d41ab0759202275a0bc031c7ccc3ca554d04f26b72e5d1a48385d9e32cc",
      "rules_applied": [
        "name-the-omission"
      ]
    },
    {
      "file": "counting-rules.md",
      "role": "The eight rules every number here is produced by, in full \u2014 including the prefix-cache trap (a naive reading is 96x wrong) and the thread-count confounder (78% on one box).",
      "media_type": "text/markdown",
      "bytes": 11432,
      "sha256": "5431b99fdbd1e9b700fdf655596fc6eb2f500366ce03f4d126b7fc08c656cb42",
      "rules_applied": []
    },
    {
      "file": "prereg.md",
      "role": "The pre-registration the run was scored against, the three machines by class, and all eight deviations from it \u2014 including the two that made a published figure worse.",
      "media_type": "text/markdown",
      "bytes": 8949,
      "sha256": "647b2eb34ab24d88498369867684514bcd845bf97b213e6eab0214d17c712eba",
      "rules_applied": [
        "name-the-omission"
      ]
    },
    {
      "file": "analysis/dispersion.md",
      "role": "The min/max behind every published decode median, 19 cells, with its spread rule stated: (max - min) divided by the MEDIAN.",
      "media_type": "text/markdown",
      "bytes": 1946,
      "sha256": "3e0760b45b6346d5e83fd74540212c88dd125e1c7f2455d99c1203ee53e926db",
      "rules_applied": [
        "state-the-rule"
      ]
    },
    {
      "file": "analysis/bandwidth.md",
      "role": "The STREAM-triad bandwidth table, its method, the honest read of the one box that does not repeat tightly, and its spread rule: (best - worst) divided by the WORST.",
      "media_type": "text/markdown",
      "bytes": 1942,
      "sha256": "475d17eab83fe3cbb7e611971747e36c289e830971622f59fcd1456d51f51b95",
      "rules_applied": [
        "state-the-rule"
      ]
    },
    {
      "file": "analysis/cause.md",
      "role": "Why Ollama chose 4 threads: the cause in llama.cpp's own default, its four receipts, the discriminating test across two differently-built hybrid CPUs, and the candidate causes ruled out.",
      "media_type": "text/markdown",
      "bytes": 4325,
      "sha256": "b809403ef0ac2a5f49c5a1874c2820811b76cea526d6e3f917533a9e913e3304",
      "rules_applied": [
        "name-the-omission"
      ]
    },
    {
      "file": "receipts/triad-cpu-server.txt",
      "role": "The server's STREAM-triad run: 64 threads, 2,289 MiB working set, 9 scored reps, best 100.69 and worst 100.36 GB/s, with the loadavg immediately before and after.",
      "media_type": "text/plain",
      "bytes": 447,
      "sha256": "572778dad367c2d235a5cf3fac50dea5d775b03d7f221eb89823642ffcaec531",
      "rules_applied": [
        "strip-uarch-literal"
      ]
    },
    {
      "file": "receipts/triad-cpu-laptop.txt",
      "role": "The laptop's STREAM-triad run: 24 threads, best 59.31 and worst 49.21 GB/s - the 20.5% spread the kit publishes un-smoothed.",
      "media_type": "text/plain",
      "bytes": 307,
      "sha256": "a79c0cc7cb33ab4c9dd21a82a58d9bdd3cc18885364c5482f1b7f10e256a0718",
      "rules_applied": [
        "strip-uarch-literal"
      ]
    },
    {
      "file": "receipts/triad-cpu-mini.txt",
      "role": "The mini's STREAM-triad run: 16 threads, best 51.12 and worst 50.85 GB/s, plus the build note explaining why its binary was cross-built.",
      "media_type": "text/plain",
      "bytes": 797,
      "sha256": "581ebef9995338a88e463fb0a160e58ec65359f03207bcf676b8b002ba1ecd62",
      "rules_applied": [
        "strip-uarch-literal",
        "box-name-to-role"
      ]
    },
    {
      "file": "receipts/coretype-cpu-laptop.txt",
      "role": "The laptop's CPUID walk: 8 performance logical CPUs, 16 efficiency, hybrid bit set, no SMT - and llama.cpp's heuristic returning 4 where the same walk without its sibling-skip returns 8.",
      "media_type": "text/plain",
      "bytes": 854,
      "sha256": "b423a5672bbf38f3485287c8b0df35effa8037cc9dd838b13d75785d6b345f39",
      "rules_applied": []
    },
    {
      "file": "receipts/coretype-cpu-mini.txt",
      "role": "The mini's CPUID walk: 8 performance logical CPUs in 4 sibling pairs, 8 efficiency, hybrid bit set - the same heuristic returning 4, by a different route.",
      "media_type": "text/plain",
      "bytes": 661,
      "sha256": "cbcf684bb3dfaa7b8df8162754920039c0b1f8f7c455a76ad6f95164aa2c00a6",
      "rules_applied": []
    },
    {
      "file": "quality/field-exam.txt",
      "role": "The 60-item field exam, all six models: per-bar scores, totals of 57-59 out of 60, every model PASS, and each one's failure modes. A floor check, not a ranker.",
      "media_type": "text/plain",
      "bytes": 1324,
      "sha256": "f5e2c522ae9ec4bef4e4fa49f148065eebec896f69869f23958af1bd1544e7a2",
      "rules_applied": [
        "absolute-path-to-name"
      ]
    },
    {
      "file": "quality/field-exam.json",
      "role": "The same exam as structured data.",
      "media_type": "application/json",
      "bytes": 1813,
      "sha256": "e9b65cb5244ae97686a491c98b2f7a4e6fa866f906d3ea05ef4465ed33e4066b",
      "rules_applied": []
    },
    {
      "file": "quality/classifier-agreement.json",
      "role": "The 374-item moderation set: per-model item agreement against the incumbent (374, 372, 371, 370, 369), split clean / hard-negative / offensive, with every disagreement's direction.",
      "media_type": "application/json",
      "bytes": 3813,
      "sha256": "deadbafa3bf30abfaf2fb9bb28b2f848de627be035b0ec131b549a638dea7ddc",
      "rules_applied": [
        "absolute-path-to-name"
      ]
    },
    {
      "file": "quality/classifier-gates.json",
      "role": "The five pre-registered moderation gates and each candidate's verdict, including the incumbent's own 2-of-329 false-positive rate.",
      "media_type": "application/json",
      "bytes": 2891,
      "sha256": "a3c2963b65ba0cc63bdeb1cc462aa6c4ce3a630cacf3256d4601d0c9b32f53e5",
      "rules_applied": [
        "absolute-path-to-name"
      ]
    },
    {
      "file": "runs/cpu-laptop-m1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7213,
      "sha256": "5d89a781539be175b03233c18d5f436912a0b546b8b126bb5d2b24d7f25cbb6f",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-laptop-m2.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7212,
      "sha256": "f9cfe163295685b00af607e581e03188effa6579772c371874b53cb751b3dde2",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-laptop-m3.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7202,
      "sha256": "5f534c14b8d933284e7ad7f2d59835c2fd411d87f84e5141ffebe8496f81aeb9",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-laptop-m4.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7297,
      "sha256": "55f2e56be4fbd9003e70ec9e916eb83563dd45e7c887d8ad295af80f978b2d4e",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-laptop-m5.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7230,
      "sha256": "74a567868bf39341db561a957c821ac4fcadca7090f7ba30ff6edeea2c9d1c0b",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-laptop-m6.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7380,
      "sha256": "4587e23ddbe98fcdfbdbe0c18a2c6baea7c43afb5ad7717c729445eaf2fccfb7",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-laptop-x1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 613,
      "sha256": "cf90208419c626673616192f75ad0ec04f55605eb7dff86c6aed80ff536b6fc9",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 6959,
      "sha256": "c7b4038cd06538255758e5d5d198a7ca5e629ec539c6dc71e84e12d0ea2491a2",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m2.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 6948,
      "sha256": "dc5aeaa0819f745e17db05bcef3fd3ac52e3517fe6a4ceaa6ed06d05be85c982",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m3.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 6945,
      "sha256": "a7d294d7cd496729a283a3d913635652b5aec6dbe68a6b2faf731c15c8604e4d",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m4.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7038,
      "sha256": "98fa6e6f1a4f89a5b9d814046f964948784d0d74eb0c5b374fcdf0306059cc1e",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m5.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 6983,
      "sha256": "85e3f79264cd4304ba0a16953365d0bedffd2a0f85451ebfbf3f04d78c17747d",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m6-memfit.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 555,
      "sha256": "63f805742b8178079a77aaeec959bcf9c386adc2f2026f8ad24562e3da47ba06",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-m6.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7125,
      "sha256": "aa5b77accf61826d1bda9382670f076d3185732a90ddd9daf019485d8060cd47",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-mini-x1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 612,
      "sha256": "ec2c88f57a79be1267cab8d87ac9f9519d1c0d0e1d6b64fd2dd307f097819749",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-server-m1.CONTAMINATED.json",
      "role": "THE ARM THE BENCH DISCARDED, kept under that name. Its nvidia snapshot is the only one in the pack that is not byte-identical before and after - a second job started on that machine's parked card partway through. Its own size_vram stayed 0. The check fired; this is what a fired check looks like.",
      "media_type": "application/json",
      "bytes": 6944,
      "sha256": "2b917336be50d38b92828e862834999798a9e6a8bdafdd70c0c871824eca36d0",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-m1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7251,
      "sha256": "2636fc14eee363de4b21aa40ca4e874aae784e8b7069a6229801c42d194ef835",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-m2.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7251,
      "sha256": "9629422bf027eb504ae590b9ace04c3bfeb3f59f1f702c4ab24a4fed0808d21e",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-m3.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7224,
      "sha256": "ab34463cea0b4feea8cfc5db57b10cdf0c33fd5b1f474975e68480c70805f84f",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-m4.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7318,
      "sha256": "976b282c87f1953df6e1d8e11e2bc2023f1e0d69501300d8a36686c3d2bde27c",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-m5-c1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 1446,
      "sha256": "47235eaf0579d7562ac20c4ec159f9449c956491dd9c7a86626b2a7c1b4c8b2e",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-server-m5-c2.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 2066,
      "sha256": "269caf0ab6ba182a98075954641b8043c2dbad7251b2e2be26089ee5f03c0055",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-server-m5.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7257,
      "sha256": "a7145efd690b7251a702bc3b62adc7a4d2f207f0d3e6371940b6ae7d86941fda",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-m6.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 7381,
      "sha256": "a596d0e1cb67fb97d1da7e5d603282cccf033b593c729a49ae8dc8bd8bc2c715",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/cpu-server-x1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 613,
      "sha256": "68f582a1883ac55ebd6807d0fcf89cedb668bfd2633c9012d970f24ce9d9ed50",
      "rules_applied": []
    },
    {
      "file": "runs/cpu-server-x2.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 5614,
      "sha256": "bfafee2cfc49f6cb64f0fe354c50e5b510e1f4b38f142a7dfd9a14e33c75f72f",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "runs/firstpass-gemma4-26b.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 1802,
      "sha256": "fb59f1a13ea62e350e9932893d12c47c14bb8e389f5c569f109047330e4244f7",
      "rules_applied": []
    },
    {
      "file": "runs/fresh-gemma4-26b-rerun.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 909,
      "sha256": "835ae3885cb11472d4f69bb8943077dd37d90ede507854e624190dcf69589c4c",
      "rules_applied": []
    },
    {
      "file": "runs/fresh-gemma4-26b.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 901,
      "sha256": "3e4d20d0806c4dafada2217ca97a299804ec8005d3a5079f010b36a27682d49e",
      "rules_applied": []
    },
    {
      "file": "runs/fresh-qwen3.5-9b.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 906,
      "sha256": "aaeb571711196696de4311586cb1502f56c490beedfa65c9666efd5157ca61b7",
      "rules_applied": []
    },
    {
      "file": "runs/threadsweep-cpu-laptop-m1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 1044,
      "sha256": "6ebb77b7dee6ea82946f162f654ed04618c814626f88ab01416849f1d5334101",
      "rules_applied": []
    },
    {
      "file": "runs/threadsweep-cpu-mini-m1.json",
      "role": "One arm of the main bench: model tag and digest, box class, UTC window, the frozen prompt's sha256, Ollama's version, the discarded warmup, every scored run's raw fields, the cold-load reps, runner peak RSS and thread count, the loadavg contention receipt, and the GPU receipt where the arm carries one.",
      "media_type": "application/json",
      "bytes": 1041,
      "sha256": "e8b7802a9c3ddf4aad224f09fa9b5acf51cfae850c557703b64fc2a45e5c735e",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/default4-cpu-mini-m1.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4913,
      "sha256": "d7db992af86b94020f10d06cde265f77f017ad71d650b59ee67702d276432c1c",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/default4-cpu-mini-m2.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4910,
      "sha256": "a59820088e9e708535e7fbcc605d573ad1581931db694e7c83f55080f436cb39",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/default4-cpu-mini-m3.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4903,
      "sha256": "251e101ec94daf4ad29a4751d65ff1e02ecbc28d72ea69dc26224105bf93d869",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/default4-cpu-mini-m4.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4955,
      "sha256": "d763f30ed78aab9f3ca27f50c2afcbd628c22d7fc3019861ea8c114f67c98411",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/default4-cpu-mini-m5.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4924,
      "sha256": "6bb50ad2ba7552fac9a8eeec5077dbee46cb0ed3b217ba06fda65cdbe7a4a26a",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/default4-cpu-mini-m6.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 5001,
      "sha256": "c69b166abc057da8789fec0644f04ac7a3286e376fcd68f85b293c7a28ca85eb",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/sweep-cpu-laptop-m1-stability.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 11882,
      "sha256": "c93404315ac6462c6979edd1109e490fa56d654c30cfa4c63fd6663db1e04e6a",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/sweep-cpu-laptop-m1.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 35051,
      "sha256": "252651a9098eb420f152ecea940f60cca5bccdaccd6c1a9b31cce529a0bb3e9a",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/sweep-cpu-mini-m1-stability.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 11939,
      "sha256": "07287f5092d9e21343fb0b69eecc82a0c2e7a216efcce1b7214a233a6f717733",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/sweep-cpu-mini-m1.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 23552,
      "sha256": "a932a0643e97952f9c0b721a459f3c281198aa4cea6d767c7d7c8f6a37dda7ae",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/sweep-cpu-mini-m6-moe.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 29839,
      "sha256": "e27b7e6a3f5b51df1e029e43c4603d779f250427060dbf874f72c4f33e1fe7d0",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/tuned-cpu-laptop-m1.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4900,
      "sha256": "d0ce0fb98c0ebbc8a3ee1a972ece98be47b4f0b60c7b45a302806e2eeb6b2561",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/tuned-cpu-laptop-m2.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4894,
      "sha256": "15d4def7573cc84310f34e73ea8dee6e7a1959e4cc9f9d10283f6878a8bfe48c",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/tuned-cpu-laptop-m3.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4885,
      "sha256": "1949a9fc515dd8cd664afa0cd1d2b4720d617a8ffd60a0ac6259c85520a7c739",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/tuned-cpu-laptop-m4.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4929,
      "sha256": "2385e4616fe39e711f34c17f11e6017bd273db2472204d2818a66479bb6b6bcd",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/tuned-cpu-laptop-m5.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4895,
      "sha256": "c7af4a67b82969c756b287ee933d0a9c86a695854ec18dbe0b9a51791d581138",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/tuned-cpu-laptop-m6.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4969,
      "sha256": "acbcb52660df65d721b17ecb6366459b73a4cdc4b913da2a533f1b7ee190588d",
      "rules_applied": [
        "local-offset-to-utc"
      ]
    },
    {
      "file": "supplement-runs/tuned-cpu-mini-m1.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4928,
      "sha256": "0d16c6e66fc586ee4d4ac36f3c17c69c79f2ca7f510976042e4c700b9e429511",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/tuned-cpu-mini-m2.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4917,
      "sha256": "812c60604e0b8042890531a595768f75e1ac519d6ead92bf34c0f6a7f5b428b1",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/tuned-cpu-mini-m3.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4917,
      "sha256": "2f2d5cce812008e1b53d106e581a988549fe6366cce6ef463773d9281a98cbb8",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/tuned-cpu-mini-m4.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4966,
      "sha256": "b383878dc267f8b7d38f10017894e52a47a8b060017296889b42ed490eff696a",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/tuned-cpu-mini-m5.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 4925,
      "sha256": "fd722d4095bf25e93d723a3c3d0feaeb2cb78e486be69b4a9610f0a12bb9a02d",
      "rules_applied": []
    },
    {
      "file": "supplement-runs/tuned-cpu-mini-m6.json",
      "role": "One arm of the evening supplement: same shape as a main-bench arm, plus arm_label and the num_thread it was pinned to.",
      "media_type": "application/json",
      "bytes": 5006,
      "sha256": "63f8aa2e206fe8de33fd2c2a9c8e6267b5b5e3388b37966774989594baf7e4e6",
      "rules_applied": []
    }
  ]
}