{
  "schema": "s2s-bench-v1",
  "kind": "data-kit-index",
  "exhibit": "the-quantization-ladder",
  "exhibit_title": "The quantization ladder",
  "exhibit_url": "https://research.strata2signal.com/the-quantization-ladder/",
  "published_utc": "2026-10-19",
  "licence": "CC BY 4.0",
  "attribution": "strata→signal research, research.strata2signal.com",
  "contact": "hello@strata2signal.com",
  "run_id": "quantization-ladder-2026-10-14",
  "run_window_utc": {
    "smoke_utc": [
      "2026-10-15T05:31:26Z",
      "2026-10-15T05:32:20Z"
    ],
    "first_sitting_utc": [
      "2026-10-15T05:34:19Z",
      "2026-10-15T06:10:47Z"
    ],
    "second_sitting_utc": [
      "2026-10-15T10:33:13Z",
      "2026-10-15T11:01:54Z"
    ],
    "reading_written_utc": "2026-10-15T11:01:54Z",
    "note": "Two sittings, from the ladder's own log: the first ran Q4_K_M and Q8_0 whole and Q6_K's speed ladder, and stopped at Q6_K's question step; the second ran Q6_K's questions again from the start, then Q5_K_M and Q3_K_M. The smoke run checked the instrument before the first sitting."
  },
  "card_class": "one GeForce RTX 3090 (24 GB) at a 300 W limit, beside a second GeForce RTX 3090 kept idle and hidden from the bench",
  "sanitisation": "Machine readers start at `index.json`, which lists every file in this kit but three, each with its byte count and its sha256; the three it leaves out are itself, `provenance.json` and this directory's index page. `provenance.json` is the other half: it names every rule applied to every file, with the count of times each one fired, and says whether any rule touched each file.",
  "files_note": "`index.json` lists every file in this kit but itself, `provenance.json` and this directory's index page, each with its sha256; `provenance.json` names every rule and how many times it fired on each file.",
  "file_count": 64,
  "files": [
    {
      "file": "AMENDMENT-1.md",
      "bytes": 6202,
      "sha256": "f232b1907c0b07ce2e4e4f5da99c4a097a321731e1cb37d38f52e2b23a1ae3c4",
      "role": "The pre-registration's one amendment, stamped 2026-10-14T22:15:53Z and 2026-10-15T02:18:44Z, before any scored call: the question step beside an idle second board (A1), the repeat gate when the repeat cannot be judged (A2), a path (A3), and the card picked by its link (A4). A public copy cut by the same rules; the digests its two stamp lines print are withheld."
    },
    {
      "file": "PREREG-QUANT.md",
      "bytes": 16631,
      "sha256": "40607a11804285f6bcacca9e8f7c62233902173c32a83b8c3b8d1252e60553af",
      "role": "The pre-registration, frozen 2026-10-14T20:52:44Z before any scored call: the question, the five builds and their manifest digests, the instruments, the order, the eight predictions with their bars, the gates and what the page may say. A public copy cut by the rules in `provenance.json`; the digest its FROZEN line prints is withheld, because it hashes the private original."
    },
    {
      "file": "READING.md",
      "bytes": 4209,
      "sha256": "a725aad9b730d6c901ddae860531a0f74710cf02681179a0d428366eb6437ea7",
      "role": "The reading, written by the bench's own reader at 2026-10-15T11:01Z after the last level: the speed table, the answers table, each prediction's verdict with the figures that decided it, and the one stop."
    },
    {
      "file": "README.md",
      "bytes": 7341,
      "sha256": "b9c11006583d70e9291291f5cb098e154a560b7f4854d429416e7a9364300676",
      "role": "The kit's guide: which file answers which table on the page, what was withheld and why."
    },
    {
      "file": "anchor/anchor-q4_K_M.json",
      "bytes": 10778,
      "sha256": "9b912475e4e8132862881fd01a2882ddf9a09d7eb96ac6f4bd581565a77ed64c",
      "role": "The 2026-09-23 Q4_K_M reading the repeat (prediction 4) was judged against: its strict choice on each of the 108 six-way items, keyed by 16 hex characters of a hash of the item's id and of its prompt. Its 48 doorman items are withheld whole (the doorman's set is private); the counts beside them are kept."
    },
    {
      "file": "logs/ladder.log",
      "bytes": 1337,
      "sha256": "a918f762d1d9f11e701fdacf2b361bc5a096f720b307478cf11e4ec7a1b6a041",
      "role": "The ladder's own log: each level's start and end in UTC, the halt of 2026-10-15T06:10:47Z and the restart at 10:33:13Z."
    },
    {
      "file": "reading.json",
      "bytes": 8611,
      "sha256": "68384f0100d29843813fbedd928f066361ef6068f54518341ba26eda99554cdc",
      "role": "The same reading as data: every figure in READING.md, per level, with each prediction's verdict and the inputs it was judged on."
    },
    {
      "file": "receipts/pull-q3_K_M.json",
      "bytes": 544,
      "sha256": "2408bd8f491c44d057c2e62ce54399958783b01851ea4a85ffce8d7b7f436287",
      "role": "A level's pull receipt: the tag, the pinned digest, the manifest's digest and size, whether the build was already in the bench's store, and the disk floor."
    },
    {
      "file": "receipts/pull-q4_K_M.json",
      "bytes": 499,
      "sha256": "eea7a328b5b678e4b6b7a3dd41a6e2a6b8bd6d0843d8bb89a39017bd33a2326e",
      "role": "A level's pull receipt: the tag, the pinned digest, the manifest's digest and size, whether the build was already in the bench's store, and the disk floor."
    },
    {
      "file": "receipts/pull-q5_K_M.json",
      "bytes": 544,
      "sha256": "e4d20ad07dd8d1f3237fc335d241b978a0e6cf133099083cbc820ee53c501a58",
      "role": "A level's pull receipt: the tag, the pinned digest, the manifest's digest and size, whether the build was already in the bench's store, and the disk floor."
    },
    {
      "file": "receipts/pull-q6_K.json",
      "bytes": 497,
      "sha256": "d4a8387f9748c3bf3d4ce326e868fac5f0a295e4a8d08814220de345674eade1",
      "role": "A level's pull receipt: the tag, the pinned digest, the manifest's digest and size, whether the build was already in the bench's store, and the disk floor."
    },
    {
      "file": "receipts/pull-q8_0.json",
      "bytes": 542,
      "sha256": "74a062cdaff856fb66a8902e95c318cb9d9d87c12e3343ed5c068ec60ad4b8b8",
      "role": "A level's pull receipt: the tag, the pinned digest, the manifest's digest and size, whether the build was already in the bench's store, and the disk floor."
    },
    {
      "file": "receipts/stage-20261015T051947Z.txt",
      "bytes": 1239,
      "sha256": "0a1de370d82a62eb6d1c72e3447ea5f753828178358c22bb3fda90b445f50db4",
      "role": "A stage receipt, read before a sitting began: the card picked and its link, its posture, the runtime, the disk, and the battery backup's load."
    },
    {
      "file": "receipts/stage-20261015T053354Z.txt",
      "bytes": 1243,
      "sha256": "02a7f6ea03c3643a745e54eddfb540e149915bf8aca9f0eb54863b7251d5c2fd",
      "role": "A stage receipt, read before a sitting began: the card picked and its link, its posture, the runtime, the disk, and the battery backup's load."
    },
    {
      "file": "receipts/stage-20261015T103251Z.txt",
      "bytes": 1236,
      "sha256": "32d928f465e0a2aa41fc5eebed23146071e61e57ef5e227f2b36aeb7b7bc3cf3",
      "role": "A stage receipt, read before a sitting began: the card picked and its link, its posture, the runtime, the disk, and the battery backup's load."
    },
    {
      "file": "receipts/stop-smoke-q4_K_M-20261015T053125Z.guard.json",
      "bytes": 1466,
      "sha256": "fab0b078e3ad76d9eab7c9943b99c0fc36ed06e5fdb05bda2bdfc2fa8f382f8d",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-speed-q3_K_M-20261015T105040Z.guard.json",
      "bytes": 1505,
      "sha256": "0a4596c7e093386986eb292fb3de973899e364807940eb988b42714ef2a155f7",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-speed-q4_K_M-20261015T053420Z.guard.json",
      "bytes": 1505,
      "sha256": "b5f067e313874d143a66997d377b3f5865e1fb0da4fec0b62bfe99d85aa6515c",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-speed-q5_K_M-20261015T103805Z.guard.json",
      "bytes": 1505,
      "sha256": "d0907e9196687ef7e0d782801234983638dd2fa0089981d07c4b8c5271519a03",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-speed-q6_K-20261015T060056Z.guard.json",
      "bytes": 1499,
      "sha256": "75229ad53dbdbbc51bc2589e20fa13d308eeebfc0c87773aa5bbc339837f558d",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-speed-q8_0-20261015T054759Z.guard.json",
      "bytes": 1499,
      "sha256": "e08b62fd168f963035f5c867e4347d0433e266f35a52a1b55c253e768436f409",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-tasks-q3_K_M-20261015T105840Z.guard.json",
      "bytes": 1105,
      "sha256": "7e7ed27b7559e57959809301b954acb4d724a21bb8ca5c42d0a2bb2e099d718d",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-tasks-q4_K_M-20261015T054143Z.guard.json",
      "bytes": 1105,
      "sha256": "3e39be9bf793f42293df3d41dcc55596870821ba87c00b6a22a80091a5cc3423",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-tasks-q5_K_M-20261015T104619Z.guard.json",
      "bytes": 1105,
      "sha256": "5af9f61e6d928a4d45b80d1413aa0aa1dec6a9b6fcbfd3186575bfa2403d5f0b",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-tasks-q6_K-20261015T061017Z.guard.json",
      "bytes": 1102,
      "sha256": "d37b3704bd42be279a53b744be1fb4ff52968cdb1d551a362e79444103c7a9e0",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-tasks-q6_K-20261015T061017Z.jsonl",
      "bytes": 132,
      "sha256": "5a675119d31ea8fe177cfbff4f8f0a4c086f7b65d34cb2f5ea5997551340b63b",
      "role": "The one stop: the card watcher's line when it stopped Q6_K's first question step at 2026-10-15T06:10:40Z."
    },
    {
      "file": "receipts/stop-tasks-q6_K-20261015T103319Z.guard.json",
      "bytes": 1099,
      "sha256": "3e48cdc19ea5d1e11b437d485e57d72d955006348db5387c88c6f9f80d387d61",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "receipts/stop-tasks-q8_0-20261015T055559Z.guard.json",
      "bytes": 1099,
      "sha256": "b4ab8c6c6aaad563b09fb4a1b7769324142c46e5a5a9a9270412de6732786b4a",
      "role": "The safety watcher's receipt for one step: the battery backup's highest load and each board's highest core temperature over the step, against the stop thresholds (80 % load, 83 °C)."
    },
    {
      "file": "results/speed/READ-gpu-ql3090-24g-nemo-q3-k-m-control-3090-q3_K_M-ladder.json",
      "bytes": 4874,
      "sha256": "d7fd7723b9e554b529a45506bffbacf34c68fceb5f8bdb94d59bb8fe6a573cf7",
      "role": "One level's speed ladder as the bench's reader summarised it: per window, whether the model was whole on the card, the median writing speed, first token, board watts, joules per 1,000 tokens, run range, spread and gates."
    },
    {
      "file": "results/speed/READ-gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-ladder.json",
      "bytes": 4872,
      "sha256": "3d6ee137352b62bbc5408b9706f8cf1ba228c0e3e97db9534edd46085b2294c5",
      "role": "One level's speed ladder as the bench's reader summarised it: per window, whether the model was whole on the card, the median writing speed, first token, board watts, joules per 1,000 tokens, run range, spread and gates."
    },
    {
      "file": "results/speed/READ-gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-smoke-ladder.json",
      "bytes": 1012,
      "sha256": "922e0b5c7b121aba6448d480e6e3cd52b6a0122f0f633be43d8f07673e47af9c",
      "role": "One level's speed ladder as the bench's reader summarised it: per window, whether the model was whole on the card, the median writing speed, first token, board watts, joules per 1,000 tokens, run range, spread and gates. This is the smoke run of 2026-10-15T05:31Z (Q4_K_M at 4,096 tokens only), which checked the instrument before the ladder and sits outside the registered order."
    },
    {
      "file": "results/speed/READ-gpu-ql3090-24g-nemo-q5-k-m-control-3090-q5_K_M-ladder.json",
      "bytes": 4876,
      "sha256": "d62c5e5b9ee83a3148d6e2a2ca6ed941b5c633d4f773199043fcd52e7e780dd8",
      "role": "One level's speed ladder as the bench's reader summarised it: per window, whether the model was whole on the card, the median writing speed, first token, board watts, joules per 1,000 tokens, run range, spread and gates."
    },
    {
      "file": "results/speed/READ-gpu-ql3090-24g-nemo-q6-k-control-3090-q6_K-ladder.json",
      "bytes": 4862,
      "sha256": "44c7e2262226eb55315a7a1af37df554dc3108473c0281abcfe7ca10832f626a",
      "role": "One level's speed ladder as the bench's reader summarised it: per window, whether the model was whole on the card, the median writing speed, first token, board watts, joules per 1,000 tokens, run range, spread and gates."
    },
    {
      "file": "results/speed/READ-gpu-ql3090-24g-nemo-q8-0-control-3090-q8_0-ladder.json",
      "bytes": 5177,
      "sha256": "6ffc0793eb4d5a3808b44b68a1038a0ef8b8785bd7715499dfa4a0da394cfdd0",
      "role": "One level's speed ladder as the bench's reader summarised it: per window, whether the model was whole on the card, the median writing speed, first token, board watts, joules per 1,000 tokens, run range, spread and gates."
    },
    {
      "file": "results/speed/gpu-ql3090-24g-nemo-q3-k-m-control-3090-q3_K_M-ladder.json",
      "bytes": 406594,
      "sha256": "af721e013666e286e9845543fc501496f5bb21f6154ccc2d11724824c91b479f",
      "role": "One level's speed ladder as the harness wrote it: every scored run's timings and token counts, the card's power and memory samples, the posture reads, the fit verdict per window and the battery backup's readout before and after."
    },
    {
      "file": "results/speed/gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-ladder.json",
      "bytes": 409279,
      "sha256": "b1281566d719e48f833bf424096d6851040a81f697ae8984a679304247e996d2",
      "role": "One level's speed ladder as the harness wrote it: every scored run's timings and token counts, the card's power and memory samples, the posture reads, the fit verdict per window and the battery backup's readout before and after."
    },
    {
      "file": "results/speed/gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-smoke-ladder.json",
      "bytes": 64115,
      "sha256": "4276d7f4919691e74028f19fe99491cff206ce9772686d732ee3e775398639d6",
      "role": "One level's speed ladder as the harness wrote it: every scored run's timings and token counts, the card's power and memory samples, the posture reads, the fit verdict per window and the battery backup's readout before and after. This is the smoke run of 2026-10-15T05:31Z (Q4_K_M at 4,096 tokens only), which checked the instrument before the ladder and sits outside the registered order."
    },
    {
      "file": "results/speed/gpu-ql3090-24g-nemo-q5-k-m-control-3090-q5_K_M-ladder.json",
      "bytes": 411138,
      "sha256": "0fe6e11a13b2f289bd8243b9be1e8586301c2be3a6af3dcbdfe1576757d81544",
      "role": "One level's speed ladder as the harness wrote it: every scored run's timings and token counts, the card's power and memory samples, the posture reads, the fit verdict per window and the battery backup's readout before and after."
    },
    {
      "file": "results/speed/gpu-ql3090-24g-nemo-q6-k-control-3090-q6_K-ladder.json",
      "bytes": 410155,
      "sha256": "79bac3339518e3bac6d6270fdce26022ba19d0f6bff335817026310c7011fadb",
      "role": "One level's speed ladder as the harness wrote it: every scored run's timings and token counts, the card's power and memory samples, the posture reads, the fit verdict per window and the battery backup's readout before and after."
    },
    {
      "file": "results/speed/gpu-ql3090-24g-nemo-q8-0-control-3090-q8_0-ladder.json",
      "bytes": 362046,
      "sha256": "520200a5fe8c5d572579c643bf7313e5751133d4767219d40d696313bff1dbf2",
      "role": "One level's speed ladder as the harness wrote it: every scored run's timings and token counts, the card's power and memory samples, the posture reads, the fit verdict per window and the battery backup's readout before and after."
    },
    {
      "file": "results/tasks/figures.json",
      "bytes": 48105,
      "sha256": "dabef85cbba3611fee7e15f8a291285f11a5a2fea87f154396bfa2cf125d390d",
      "role": "The question step's figures for every level: the six-way and the doorman's set scored strict and lenient, board and wall energy, timings, the probes and the fit on the named card. The doorman's set appears as counts only."
    },
    {
      "file": "results/tasks/q-q3_K_M/arm.json",
      "bytes": 34055,
      "sha256": "a652b5365cf6c46bd36b9e0323726ebb0fc9cf50e926a4f745240da2c31fed63",
      "role": "One level's question step as the instrument wrote it: the model and its file header (parameter count, context length), the card's posture before and after, the fit and growth on the named card, the probes' summary and the step's start and end in UTC."
    },
    {
      "file": "results/tasks/q-q3_K_M/probes.json",
      "bytes": 8818,
      "sha256": "28abc242c2a282507ab08799613064eed3265ec0c363f9d6bd3a462267ff6b84",
      "role": "One level's probes: writing 256 tokens three times, reading a long prompt, and the first token on 12 six-way prompts, with the board's energy over each."
    },
    {
      "file": "results/tasks/q-q3_K_M/q-q3_K_M.c.jsonl",
      "bytes": 119350,
      "sha256": "fa57a6f72b8d576703ede6778145eb3344446e47c503f3eda2f0b18c8780bf42",
      "role": "One level's six-way rows: one line per item (108), with the reply, the parse, the choice, the server's own timings and the card's posture."
    },
    {
      "file": "results/tasks/q-q3_K_M/q-q3_K_M.c.report.json",
      "bytes": 5845,
      "sha256": "c4f543a69b4b60ba1ab4ec73d02e1d66a8716a4225de19bbb50134bb66b84570",
      "role": "One level's six-way report as the instrument wrote it: the score, the energy window and the idle reading, and the instrument's settings."
    },
    {
      "file": "results/tasks/q-q4_K_M/arm.json",
      "bytes": 34336,
      "sha256": "2470ebbbee5145a56fbc11c98cf9a85a273fec6840b6748059135b75e6da31b5",
      "role": "One level's question step as the instrument wrote it: the model and its file header (parameter count, context length), the card's posture before and after, the fit and growth on the named card, the probes' summary and the step's start and end in UTC."
    },
    {
      "file": "results/tasks/q-q4_K_M/probes.json",
      "bytes": 9133,
      "sha256": "d8196bcd0fa182e82fe19859555b79483a298a5458177d40d919f476517bdbd0",
      "role": "One level's probes: writing 256 tokens three times, reading a long prompt, and the first token on 12 six-way prompts, with the board's energy over each."
    },
    {
      "file": "results/tasks/q-q4_K_M/q-q4_K_M.c.jsonl",
      "bytes": 119260,
      "sha256": "1baa5cf4398fb0079ed573f9788c737331fa18b48fe255f76fc843ae2673c173",
      "role": "One level's six-way rows: one line per item (108), with the reply, the parse, the choice, the server's own timings and the card's posture."
    },
    {
      "file": "results/tasks/q-q4_K_M/q-q4_K_M.c.report.json",
      "bytes": 5846,
      "sha256": "4b09fa2cba0b853cbf825cd3b51032db028e3961f9645fdf0b98ab166e7a5189",
      "role": "One level's six-way report as the instrument wrote it: the score, the energy window and the idle reading, and the instrument's settings."
    },
    {
      "file": "results/tasks/q-q5_K_M/arm.json",
      "bytes": 34016,
      "sha256": "125ae897da17f69427f70c0d7dd39cf15b5e450e2adb15e48ae829e6bc49d526",
      "role": "One level's question step as the instrument wrote it: the model and its file header (parameter count, context length), the card's posture before and after, the fit and growth on the named card, the probes' summary and the step's start and end in UTC."
    },
    {
      "file": "results/tasks/q-q5_K_M/probes.json",
      "bytes": 8812,
      "sha256": "df93475b56db169fc47de17a86db38dbff2b78bc0a254c9d097fd12146931d43",
      "role": "One level's probes: writing 256 tokens three times, reading a long prompt, and the first token on 12 six-way prompts, with the board's energy over each."
    },
    {
      "file": "results/tasks/q-q5_K_M/q-q5_K_M.c.jsonl",
      "bytes": 119346,
      "sha256": "78e96892a4815d10492543e5c6be7924956de8f618f57d9e0cf2efc63978487a",
      "role": "One level's six-way rows: one line per item (108), with the reply, the parse, the choice, the server's own timings and the card's posture."
    },
    {
      "file": "results/tasks/q-q5_K_M/q-q5_K_M.c.report.json",
      "bytes": 5842,
      "sha256": "5cbd2f72c81996f7884e1cd87a7c85381af30d9a28e285895dabed76ffc3f697",
      "role": "One level's six-way report as the instrument wrote it: the score, the energy window and the idle reading, and the instrument's settings."
    },
    {
      "file": "results/tasks/q-q6_K/arm.json",
      "bytes": 34020,
      "sha256": "11ec943e424179cc3a8083f65c2309d0c1fc7e2083e8128361602608877ad14e",
      "role": "One level's question step as the instrument wrote it: the model and its file header (parameter count, context length), the card's posture before and after, the fit and growth on the named card, the probes' summary and the step's start and end in UTC."
    },
    {
      "file": "results/tasks/q-q6_K/probes.json",
      "bytes": 8813,
      "sha256": "85aa133163e6d5d2a92896e53ca68f87590a66cb915a563551225b139db94506",
      "role": "One level's probes: writing 256 tokens three times, reading a long prompt, and the first token on 12 six-way prompts, with the board's energy over each."
    },
    {
      "file": "results/tasks/q-q6_K/q-q6_K.c.jsonl",
      "bytes": 118726,
      "sha256": "4e1e1e3b36064846478450b0d263c203a33beb431d2a9bb57b34095494abce15",
      "role": "One level's six-way rows: one line per item (108), with the reply, the parse, the choice, the server's own timings and the card's posture."
    },
    {
      "file": "results/tasks/q-q6_K/q-q6_K.c.report.json",
      "bytes": 5836,
      "sha256": "af6445a4c3ea65d9f4469f50cce7cd6b2bcebe0055a3f59fe6d56ca3dc08de12",
      "role": "One level's six-way report as the instrument wrote it: the score, the energy window and the idle reading, and the instrument's settings."
    },
    {
      "file": "results/tasks/q-q8_0/arm.json",
      "bytes": 34315,
      "sha256": "7a7622211331ba93694a33b52a9764c84b1556cd3f556fe301e835354203ab7a",
      "role": "One level's question step as the instrument wrote it: the model and its file header (parameter count, context length), the card's posture before and after, the fit and growth on the named card, the probes' summary and the step's start and end in UTC."
    },
    {
      "file": "results/tasks/q-q8_0/probes.json",
      "bytes": 9131,
      "sha256": "33f1ae330e75f613b873eb38f036fb917621f35b0608069ef59165f620280f35",
      "role": "One level's probes: writing 256 tokens three times, reading a long prompt, and the first token on 12 six-way prompts, with the board's energy over each."
    },
    {
      "file": "results/tasks/q-q8_0/q-q8_0.c.jsonl",
      "bytes": 118617,
      "sha256": "d6447e9587cd1ab609fceab06f44b58ae530868ee439d12f6dd5748eeeab31de",
      "role": "One level's six-way rows: one line per item (108), with the reply, the parse, the choice, the server's own timings and the card's posture."
    },
    {
      "file": "results/tasks/q-q8_0/q-q8_0.c.report.json",
      "bytes": 5780,
      "sha256": "9a2a12f2893a75af3ee9ad52a595f9b131f6cc0b8ec4f4d033e2bb077f4b3ee1",
      "role": "One level's six-way report as the instrument wrote it: the score, the energy window and the idle reading, and the instrument's settings."
    }
  ]
}