{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "the-quantization-ladder",
 "title": "The quantization ladder",
 "dek": "We ran one model, Mistral NeMo 12B Instruct, at five of the quantization levels it is downloaded in, from Q8_0 down to Q3_K_M, each whole on one 24 GB card at 300 W and measured the same way, and every step down to Q4_K_M wrote faster with no change in answers the paired test could tell from Q8_0, while Q3_K_M wrote slower than Q4_K_M and lost 16 answers Q8_0 got right.",
 "published": "2026-10-19",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0",
 "status": "pending-judge",
 "pack_note": "no judge is seated, so every chip is held",
 "notice": "items with published:false are held and are not this page's published words",
 "source": {
  "url": "https://research.strata2signal.com/the-quantization-ladder/",
  "md_url": "https://research.strata2signal.com/the-quantization-ladder/index.md",
  "md_sha": "1c780e0ab9cff5b6c7762c442025c632d76f96df23e10d5804fbe05555a32464",
  "html_sha": "227a691ce00ca20f5b53a37f218dabe5f5208d0ff7a1cc05ad89f3a6b101203e",
  "anchors_sha": "d036825b8ce3de414008ebc6d5407600f0e88cbe01cb17c13d91f08d448b3a24"
 },
 "short": {
  "paragraph": "Mistral NeMo 12B Instruct at five quantization levels, downloads from 5.67 GiB (Q3KM) to 12.13 GiB (Q80), each run whole on one GeForce RTX 3090 at 300 W with the questions held fixed. At a 4,096-token window it wrote 60.5 tokens a second at Q80, rising at every step down to 88.2 at Q4KM, then fell back to 74.9 at Q3KM: the smallest file was slower than the one above it. With the cache at 8 bits, Q80 held a window of 98,304 tokens whole; the other four held 131,072, the 128K window Mistral AI advertises. On 108 frozen questions the levels scored from 38 (Q3KM) to 57 (Q5KM). The paired test could not tell any level from Q4KM to Q6K from Q80 (Q4KM came nearest, p 0.0654), but it could tell Q3KM (p under 0.0001). On the 48 held-out lines every level scored 33 or 34. Of the eight predictions, five held and three failed; none was left undecided.",
  "from_dek": false,
  "counts": {
   "words": 5145,
   "minutes": 23,
   "tables": 6,
   "kit": true
  },
  "bullets": [
   {
    "text": "the registry's listings were read on 2026-10-14",
    "figure": "2026-10-14",
    "cite": "the-question",
    "quote": "The five builds come from one publisher, and the registry's listings (read 2026-10-14 at 20:47 UTC) show the same chat template, licence and settings files in all five, byte for byte.",
    "span": [
     3150,
     3161
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-question",
   "heading": "The question",
   "level": 2,
   "span": [
    2892,
    3327
   ],
   "chunks": [
    [
     2892,
     3327
    ]
   ],
   "chars": 435,
   "digest": "the article asks what each step down in bits costs in answers and what it buys in speed and memory for one model on one card, while holding everything else still.",
   "digest_skipped": null
  },
  {
   "id": "why-it-matters",
   "heading": "Why it matters if you are choosing a download",
   "level": 2,
   "span": [
    3327,
    4219
   ],
   "chunks": [
    [
     3327,
     4219
    ]
   ],
   "chars": 892,
   "digest": "choosing a download level decides whether a model fits on a card with room for conversation, how fast it writes, and how many answers change.",
   "digest_skipped": null
  },
  {
   "id": "the-five-files",
   "heading": "The five files",
   "level": 2,
   "span": [
    4219,
    6453
   ],
   "chunks": [
    [
     4219,
     6453
    ]
   ],
   "chars": 2234,
   "digest": "five builds of mistral-nemo:12b-instruct-2407 from the Ollama library are examined, with each build pinned by its manifest digest and checked against it.",
   "digest_skipped": null
  },
  {
   "id": "how-it-was-measured",
   "heading": "How it was measured",
   "level": 2,
   "span": [
    6453,
    10216
   ],
   "chunks": [
    [
     6453,
     10216
    ]
   ],
   "chars": 3763,
   "digest": "measurements used a GeForce RTX 3090 at a 300 W limit, Ollama 0.32.13, a speed ladder, 108 frozen questions, 48 held-out lines, and various probes.",
   "digest_skipped": null
  },
  {
   "id": "what-we-predicted",
   "heading": "What we predicted, before the first scored call",
   "level": 2,
   "span": [
    10216,
    12614
   ],
   "chunks": [
    [
     10216,
     12614
    ]
   ],
   "chars": 2398,
   "digest": "eight predictions were made regarding speed, the bytes law, window size, a repeat of a previous reading, answers, the doorman, short calls, and memory.",
   "digest_skipped": null
  },
  {
   "id": "what-voids-a-reading",
   "heading": "What voids a reading",
   "level": 2,
   "span": [
    12614,
    13836
   ],
   "chunks": [
    [
     12614,
     13836
    ]
   ],
   "chars": 1222,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-first-measured-row",
   "heading": "The first measured row",
   "level": 2,
   "span": [
    13836,
    15036
   ],
   "chunks": [
    [
     13836,
     15036
    ]
   ],
   "chars": 1200,
   "digest": "a smoke run checked the instrument with Q4_K_M at a 4,096-token window, writing at 88.548 tokens a second.",
   "digest_skipped": null
  },
  {
   "id": "when-it-ran",
   "heading": "When it ran, and the one stop",
   "level": 2,
   "span": [
    15036,
    16630
   ],
   "chunks": [
    [
     15036,
     16630
    ]
   ],
   "chars": 1594,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "speed-and-memory",
   "heading": "What each step down bought: speed and memory",
   "level": 2,
   "span": [
    16630,
    20152
   ],
   "chunks": [
    [
     16630,
     20152
    ]
   ],
   "chars": 3522,
   "digest": "each level's writing speed, first token time, board power, energy, and largest window are recorded, alongside memory growth relative to download size.",
   "digest_skipped": null
  },
  {
   "id": "answers",
   "heading": "What each step down cost: answers",
   "level": 2,
   "span": [
    20152,
    23977
   ],
   "chunks": [
    [
     20152,
     23977
    ]
   ],
   "chars": 3825,
   "digest": "the six-way and doorman's set scores are provided for each level, including McNemar tests, Wilson 95 per cent intervals, and energy per correct answer.",
   "digest_skipped": null
  },
  {
   "id": "the-predictions-judged",
   "heading": "The predictions, judged",
   "level": 2,
   "span": [
    23977,
    26051
   ],
   "chunks": [
    [
     23977,
     26051
    ]
   ],
   "chars": 2074,
   "digest": "the eight frozen predictions are judged with verdicts and the deciding figures from the measurements.",
   "digest_skipped": null
  },
  {
   "id": "what-it-means",
   "heading": "What it means if you are choosing a download",
   "level": 2,
   "span": [
    26051,
    27101
   ],
   "chunks": [
    [
     26051,
     27101
    ]
   ],
   "chars": 1050,
   "digest": "for this model on a 24 GB GeForce RTX 3090, Q5_K_M is suggested as a pick, while Q4_K_M is the speed pick and Q3_K_M is the level to skip.",
   "digest_skipped": null
  },
  {
   "id": "what-this-page-cannot-say",
   "heading": "What this page cannot say",
   "level": 2,
   "span": [
    27101,
    28265
   ],
   "chunks": [
    [
     27101,
     28265
    ]
   ],
   "chars": 1164,
   "digest": "the results do not carry to other models, publishers, runtimes, or cards, and do not measure perplexity, long answers, code, or sampled output.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    28265,
    29268
   ],
   "chunks": [
    [
     28265,
     29268
    ]
   ],
   "chars": 1003,
   "digest": "summary points cover speed, window, answers, and the recommended pick for this model on this card.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    29268,
    30791
   ],
   "chunks": [
    [
     29268,
     30791
    ]
   ],
   "chars": 1523,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    30791,
    31638
   ],
   "chunks": [
    [
     30791,
     31638
    ]
   ],
   "chars": 847,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    31638,
    34125
   ],
   "chunks": [
    [
     31638,
     34125
    ]
   ],
   "chars": 2487,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-f80cc541",
   "q": "what variables remain constant during the quantization comparison?",
   "a": "The five builds share the same publisher, chat template, licence, and settings files. Only the weights change, while the card, runtime, and questions remain frozen.",
   "cites": [
    "the-question"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-aafdb8b7",
   "q": "how does quantization level affect model performance and resource use?",
   "a": "The chosen level determines if the model fits on a card with room for conversation, how fast it writes tokens, and how many answers change.",
   "cites": [
    "why-it-matters"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-b6a9eb92",
   "q": "how many bits per weight does the q3_k_m build carry?",
   "a": "By the whole-file rule, the q3_k_m file carries 3.97 bits a weight.",
   "cites": [
    "the-five-files"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-285cebdc",
   "q": "what method determines the accuracy of the six-way scoring?",
   "a": "Scoring is strict, requiring the reply to be at most 8 tokens and open with the correct letter. A McNemar test is used to determine if differences in accuracy between levels are statistically significant.",
   "cites": [
    "how-it-was-measured"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-e33e06fd",
   "q": "what was the prediction regarding writing speed for q8_0 to q4_k_m?",
   "a": "The prediction stated that writing speed at 4,096 tokens rises at every step down from q8_0 to q4_k_m, with each step's gain larger than the run-to-run range.",
   "cites": [
    "what-we-predicted"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-bb708503",
   "q": "what was the writing speed of the q4_k_m smoke run?",
   "a": "The q4_k_m smoke run measured 88.548 tokens a second at a 4,096-token window.",
   "cites": [
    "the-first-measured-row"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "two-used-3080s-priced",
   "why": "shares ground with § The question · § What this page does not measure"
  },
  {
   "slug": "flash-next-why-the-first-word-comes-sooner",
   "why": "shares ground with § The question · § The question, and the instruments"
  },
  {
   "slug": "the-same-eighteen-pictures-card-by-card",
   "why": "shares ground with § How it was measured · § The EVGA 3090 Ti FTW3 Ultra's column, at 350 W and 300 W (added 2026-09-26)"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nMistral AI and NVIDIA (Mistral NeMo 12B Instruct, Apache 2.0); Ollama (the runtime, MIT, and the library serving the five builds); Georgi Gerganov and the ggml authors (GGUF and its quantization types, MIT); and the long table's six guests, whose borrowed voices make the six-way. None of them owed us anything. A small (human) team approved the bench, set the card's power limit and started every run; a fleet of AI agents chose the model, wrote the pre-registration and the kit, read the rows and drafted this page.\n\nRead first-hand on 2026-10-16 (UTC):\n\n- Ollama: the library's [tags page for `mistral-nemo`](https://ollama.com/library/mistral-nemo/tags) (the five tags, each with the first 12 characters of its digest and its size, matching the table of five); the five builds' manifests and the shared template, licence and settings files were read from the registry by the bench on 2026-10-14 at 20:47 UTC\n- Mistral AI: [Mistral NeMo](https://mistral.ai/news/mistral-nemo), 2024-07-18 (\"a 12B model built in collaboration with NVIDIA\", a context window \"of up to 128k tokens\", released \"under the Apache 2.0 license\")\n- This workshop's own released pages: [Inside a GGUF file](https://research.strata2signal.com/inside-a-gguf-file/) (the block arithmetic, read there from llama.cpp's source at commit `efe3626b`), [The compressed photograph](https://research.strata2signal.com/the-compressed-photograph/) (its two tests and the gap below four bits), [A short history of Mistral](https://research.strata2signal.com/a-short-history-of-mistral/) (NeMo's date, licence and window), [A laptop, asked the desktop's questions](https://research.strata2signal.com/a-laptop-asked-the-desktops-questions/) (the speed ladder), [Six graphics cards, the same questions](https://research.strata2signal.com/six-cards-the-same-questions/), [Reading the answer instead of writing it](https://research.strata2signal.com/reading-the-answer/) (the six-way and the doorman's set, and the doorman's three kinds of line), and [The cost of intelligence](https://research.strata2signal.com/the-cost-of-intelligence/) (the home-team flag; its kit's series carries the 2026-09-23 Q4_K_M rows the repeat is judged against)\n\nCorrections and later measurements will be added below, each dated (UTC), with a window at both ends where one applies, saying in plain words what it counts.\n\n*Licence: CC BY 4.0, the whole page — name the source and link to it.*",
 "kit": {
  "url": "https://research.strata2signal.com/the-quantization-ladder/data/",
  "licence": "CC BY 4.0",
  "files": [
   "AMENDMENT-1.md",
   "PREREG-QUANT.md",
   "READING.md",
   "README.md",
   "anchor/anchor-q4_K_M.json",
   "logs/ladder.log",
   "reading.json",
   "receipts/pull-q3_K_M.json",
   "receipts/pull-q4_K_M.json",
   "receipts/pull-q5_K_M.json",
   "receipts/pull-q6_K.json",
   "receipts/pull-q8_0.json",
   "receipts/stage-20261015T051947Z.txt",
   "receipts/stage-20261015T053354Z.txt",
   "receipts/stage-20261015T103251Z.txt",
   "receipts/stop-smoke-q4_K_M-20261015T053125Z.guard.json",
   "receipts/stop-speed-q3_K_M-20261015T105040Z.guard.json",
   "receipts/stop-speed-q4_K_M-20261015T053420Z.guard.json",
   "receipts/stop-speed-q5_K_M-20261015T103805Z.guard.json",
   "receipts/stop-speed-q6_K-20261015T060056Z.guard.json",
   "receipts/stop-speed-q8_0-20261015T054759Z.guard.json",
   "receipts/stop-tasks-q3_K_M-20261015T105840Z.guard.json",
   "receipts/stop-tasks-q4_K_M-20261015T054143Z.guard.json",
   "receipts/stop-tasks-q5_K_M-20261015T104619Z.guard.json",
   "receipts/stop-tasks-q6_K-20261015T061017Z.guard.json",
   "receipts/stop-tasks-q6_K-20261015T061017Z.jsonl",
   "receipts/stop-tasks-q6_K-20261015T103319Z.guard.json",
   "receipts/stop-tasks-q8_0-20261015T055559Z.guard.json",
   "results/speed/READ-gpu-ql3090-24g-nemo-q3-k-m-control-3090-q3_K_M-ladder.json",
   "results/speed/READ-gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-ladder.json",
   "results/speed/READ-gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-smoke-ladder.json",
   "results/speed/READ-gpu-ql3090-24g-nemo-q5-k-m-control-3090-q5_K_M-ladder.json",
   "results/speed/READ-gpu-ql3090-24g-nemo-q6-k-control-3090-q6_K-ladder.json",
   "results/speed/READ-gpu-ql3090-24g-nemo-q8-0-control-3090-q8_0-ladder.json",
   "results/speed/gpu-ql3090-24g-nemo-q3-k-m-control-3090-q3_K_M-ladder.json",
   "results/speed/gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-ladder.json",
   "results/speed/gpu-ql3090-24g-nemo-q4-k-m-control-3090-q4_K_M-smoke-ladder.json",
   "results/speed/gpu-ql3090-24g-nemo-q5-k-m-control-3090-q5_K_M-ladder.json",
   "results/speed/gpu-ql3090-24g-nemo-q6-k-control-3090-q6_K-ladder.json",
   "results/speed/gpu-ql3090-24g-nemo-q8-0-control-3090-q8_0-ladder.json",
   "results/tasks/figures.json",
   "results/tasks/q-q3_K_M/arm.json",
   "results/tasks/q-q3_K_M/probes.json",
   "results/tasks/q-q3_K_M/q-q3_K_M.c.jsonl",
   "results/tasks/q-q3_K_M/q-q3_K_M.c.report.json",
   "results/tasks/q-q4_K_M/arm.json",
   "results/tasks/q-q4_K_M/probes.json",
   "results/tasks/q-q4_K_M/q-q4_K_M.c.jsonl",
   "results/tasks/q-q4_K_M/q-q4_K_M.c.report.json",
   "results/tasks/q-q5_K_M/arm.json",
   "results/tasks/q-q5_K_M/probes.json",
   "results/tasks/q-q5_K_M/q-q5_K_M.c.jsonl",
   "results/tasks/q-q5_K_M/q-q5_K_M.c.report.json",
   "results/tasks/q-q6_K/arm.json",
   "results/tasks/q-q6_K/probes.json",
   "results/tasks/q-q6_K/q-q6_K.c.jsonl",
   "results/tasks/q-q6_K/q-q6_K.c.report.json",
   "results/tasks/q-q8_0/arm.json",
   "results/tasks/q-q8_0/probes.json",
   "results/tasks/q-q8_0/q-q8_0.c.jsonl",
   "results/tasks/q-q8_0/q-q8_0.c.report.json"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 1,
  "digests_written": 14,
  "digests_kept": 12,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 0,
   "length": 1,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 5
  },
  "new_noun_tokens_checked": [
   "0.32.13",
   "06",
   "096",
   "096-token",
   "10",
   "108",
   "128k",
   "12b-instruct-2407",
   "2026-10-14",
   "2026-10-15",
   "24",
   "3.97",
   "30",
   "300",
   "304",
   "3090",
   "4",
   "40",
   "48",
   "8",
   "88.548",
   "95",
   "98",
   "GB",
   "GeForce",
   "McNemar",
   "Ollama",
   "Q3KM",
   "Q4KM",
   "Q5KM",
   "Q80",
   "RTX",
   "UTC",
   "W",
   "Wilson",
   "q3km",
   "q4km",
   "q5km",
   "q80"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "bullet",
    "reason": "length",
    "item": "the board power was read twice a second in the speed ladder and once a second during the questions at 300 W",
    "detail": "22 words"
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 49,
    "text_sha": "e46852179b58",
    "figure": "98,304",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-10-17",
    "why": "the hand-read against the release source, 2026-10-17 (UTC): Q8_0's largest window held whole is 98,304 tokens with the cache at 8 bits; the line drops both the unit and that condition (at 16 bits the cache takes nearly twice the room)",
    "detail": "removed from this pack by hand on 2026-10-17 (UTC), after the writer ran and after the gates had counted: the hand-read against the release source, 2026-10-17 (UTC): Q8_0's largest window held whole is 98,304 tokens with the cache at 8 bits; the line drops both the unit and that condition (at 16 bits the cache takes nearly twice the room). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 133,
    "text_sha": "16d4ad3127ef",
    "section": "what-voids-a-reading",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-17",
    "why": "the hand-read against the release source, 2026-10-17 (UTC): the digest says a failed repeat, throttling or spread voids a reading, where the section labels them (instrument drift, throttled, a printed spread) and voids only what a gate names",
    "detail": "removed from this pack by hand on 2026-10-17 (UTC), after the writer ran and after the gates had counted: the hand-read against the release source, 2026-10-17 (UTC): the digest says a failed repeat, throttling or spread voids a reading, where the section labels them (instrument drift, throttled, a printed spread) and voids only what a gate names. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 111,
    "text_sha": "ee4b2599afab",
    "section": "when-it-ran",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-10-17",
    "why": "the hand-read against the release source, 2026-10-17 (UTC): the digest puts the one stop down to a foreign process, where the section says the watcher counted the bench's own short-lived runner as foreign, almost certainly, and fixed the watcher",
    "detail": "removed from this pack by hand on 2026-10-17 (UTC), after the writer ran and after the gates had counted: the hand-read against the release source, 2026-10-17 (UTC): the digest puts the one stop down to a foreign process, where the section says the watcher counted the bench's own short-lived runner as foreign, almost certainly, and fixed the watcher. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 240,
    "text_sha": "38aa72d1ac22",
    "chip_id": "c-c045203b",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-17",
    "why": "the hand-read against the release source, 2026-10-17 (UTC): the answer says a throttled card or a model that does not fit whole voids a reading, where the section flags a throttled run and marks a level that does not fit as not tested",
    "detail": "removed from this pack by hand on 2026-10-17 (UTC), after the writer ran and after the gates had counted: the hand-read against the release source, 2026-10-17 (UTC): the answer says a throttled card or a model that does not fit whole voids a reading, where the section flags a throttled run and marks a level that does not fit as not tested. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 262,
    "text_sha": "8e41e5b708aa",
    "chip_id": "c-b1a515cb",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-10-17",
    "why": "the hand-read against the v2 release source, 2026-10-17 (UTC): the answer says Q5_K_M held the advertised 128K window and drops its condition, the cache at 8 bits (at the runtime's default 16-bit cache the window does not hold), the same condition the release audit's I-3 put back on the page's short version and its pick",
    "detail": "removed from this pack by hand on 2026-10-17 (UTC), after the writer ran and after the gates had counted: the hand-read against the v2 release source, 2026-10-17 (UTC): the answer says Q5_K_M held the advertised 128K window and drops its condition, the cache at 8 bits (at the runtime's default 16-bit cache the window does not hold), the same condition the release audit's I-3 put back on the page's short version and its pick. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-10-17T02:32:16Z"
}