{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "mixture-of-experts-in-plain-words",
 "title": "Mixture of experts, in plain words: why the big number decides the fit and the small one the speed",
 "dek": "A model card's two numbers do two jobs: the big one decides whether a mixture of experts fits on your card, the small one roughly how fast it writes once it does; following Everyone on the payroll, three at the table, which taught the two numbers on models that fit, this guide adds what happens when one does not: the two ways the experts spill into system memory, shown on one model that fits whole on a 96 GB card and one that cannot.",
 "published": "2026-11-05",
 "series": [
  "guide"
 ],
 "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/mixture-of-experts-in-plain-words/",
  "md_url": "https://research.strata2signal.com/mixture-of-experts-in-plain-words/index.md",
  "md_sha": "24fbc041ef7eb0177184dc4ad284497b501ebf9a67bbc90e213dfcb553325deb",
  "html_sha": "7c424a864a1948cf3c55a140c0ab370f8988ac0e65f428eaa43af960144d8572",
  "anchors_sha": "11e2cf81dd36ff96710cc75397d83b215736e6c0470394f7a70599d07b497259"
 },
 "short": {
  "paragraph": "A mixture of experts is a model whose layers each hold many small sub-networks, the experts, and a router that picks a few of them for every token. The big number on the card counts every expert; the small number counts the ones a token meets. Any token may call any expert, so the big number decides whether the model fits on a card. Writing hauls only the weights a token needs, so once it fits, the small number roughly sets the writing speed: on one NVIDIA GeForce RTX 3090 (24 GB), at one cap and with one prompt, a 26B mixture wrote 3.7 times as fast as a dense 24B. When the experts do not fit, an engine keeps some or all of them in the computer's own memory, and the processor and that memory set the pace: llama.cpp wrote 22.490 tokens a second on an EVGA GeForce RTX 3090 Ti FTW3 Ultra (24 GB) with the experts of 34 of 48 layers in system memory, its eight processor threads setting the pace, and 102.889 on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition (96 GB) in the same desktop with every expert on the card.",
  "from_dek": false,
  "counts": {
   "words": 4099,
   "minutes": 19,
   "tables": 5,
   "kit": false
  },
  "bullets": [
   {
    "text": "the desktop's system memory has a theoretical speed of about 204.8 GB/s",
    "figure": "204.8 GB/s",
    "cite": "two-ways-to-spill",
    "quote": "Where an expert sits Published or rated speed Source The 96 GB card's own memory 1,792 GB/s NVIDIA's product page, \"1792 GB/sec\", read 2026-10-29 (UTC) The desktop's system memory, 128 GB of DDR4-3200 in eight channels About 204.8 GB/s Arithmetic: 3,200 MT/s × 8 bytes × 8 channels, theoretical, never measured The laptop's system memory, 64 GB of DDR5-5600 in two channels About 89.6 GB/s Arithmetic: 5,600 × 8 × 2, theoretical, never measured The laptop's card link 28.7 GB/s The Strata engine's own probe, on the three-card page By these ratings an expert in the desktop's system memory is reached about 9 times slower than one on its 96 GB card (204.8 against 1,792 GB/s), or computed by the processor instead: a bound within one computer, not a measurement; the laptop's memory is rated at under half the desktop's.",
    "span": [
     17857,
     17868
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-words-this-page-leans-on",
   "heading": "The words this page leans on",
   "level": 2,
   "span": [
    2585,
    4104
   ],
   "chunks": [
    [
     2585,
     4104
    ]
   ],
   "chars": 1519,
   "digest": "this section defines technical terms including a token, a parameter, a layer, an expert, and a router. it explains active parameters, card and system memory, a spill, bits per weight, and the difference between GB and GiB.",
   "digest_skipped": null
  },
  {
   "id": "two-numbers-on-every-model-card",
   "heading": "Two numbers on every model card",
   "level": 2,
   "span": [
    4104,
    6184
   ],
   "chunks": [
    [
     4104,
     6184
    ]
   ],
   "chars": 2080,
   "digest": "the text compares parameter counts and expert configurations for Qwen3.8-Flash-Next, DeepSeek V4.1 Flash, and Mixtral 8x7B. it notes that active counts are a vendor convention rather than a file field and provides a comparison table of their architectures.",
   "digest_skipped": null
  },
  {
   "id": "what-an-expert-is-and-who-picks",
   "heading": "What an expert is, and who picks",
   "level": 2,
   "span": [
    6184,
    9470
   ],
   "chunks": [
    [
     6184,
     9470
    ]
   ],
   "chars": 3286,
   "digest": "experts are narrow sub-networks within a layer, selected by a router. the section details the architecture of Qwen3.8-Flash-Next and DeepSeek V4.1 Flash, explaining how routers score experts and how the resulting outputs are weighted and summed.",
   "digest_skipped": null
  },
  {
   "id": "where-the-idea-came-from-in-four-lines",
   "heading": "Where the idea came from, in four lines",
   "level": 2,
   "span": [
    9470,
    10772
   ],
   "chunks": [
    [
     9470,
     10772
    ]
   ],
   "chars": 1302,
   "digest": "this section traces the history of mixture-of-experts from 1991 to 2024, citing the development of adaptive mixtures, sparsely-gated layers, the Switch Transformer, and the release of Mixtral 8x7B and DeepSeekMoE.",
   "digest_skipped": null
  },
  {
   "id": "why-the-big-number-decides-whether-it-fits",
   "heading": "Why the big number decides whether it fits",
   "level": 2,
   "span": [
    10772,
    11991
   ],
   "chunks": [
    [
     10772,
     11991
    ]
   ],
   "chars": 1219,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "why-the-small-number-decides-how-fast-it-writes-once-it-fits",
   "heading": "Why the small number decides how fast it writes, once it fits",
   "level": 2,
   "span": [
    11991,
    13624
   ],
   "chunks": [
    [
     11991,
     13624
    ]
   ],
   "chars": 1633,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "fits-or-spills-on-one-96-gb-card",
   "heading": "Fits, or spills, on one 96 GB card",
   "level": 2,
   "span": [
    13624,
    16331
   ],
   "chunks": [
    [
     13624,
     16331
    ]
   ],
   "chars": 2707,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "two-ways-to-spill",
   "heading": "Two ways to spill",
   "level": 2,
   "span": [
    16331,
    21764
   ],
   "chunks": [
    [
     16331,
     21764
    ]
   ],
   "chars": 5433,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-it-means-for-your-own-card",
   "heading": "What it means for your own card",
   "level": 2,
   "span": [
    21764,
    23524
   ],
   "chunks": [
    [
     21764,
     23524
    ]
   ],
   "chars": 1760,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-this-cannot-say",
   "heading": "What this cannot say",
   "level": 2,
   "span": [
    23524,
    24385
   ],
   "chunks": [
    [
     23524,
     24385
    ]
   ],
   "chars": 861,
   "digest": "the article clarifies it does not provide measurements for DeepSeek V4.1 Flash, does not establish ratios between cards, does not explain how makers' active counts are made, and does not comment on quality.",
   "digest_skipped": null
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    24385,
    25217
   ],
   "chunks": [
    [
     24385,
     25217
    ]
   ],
   "chars": 832,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "sources",
   "heading": "Sources",
   "level": 2,
   "span": [
    25217,
    29113
   ],
   "chunks": [
    [
     25217,
     29113
    ]
   ],
   "chars": 3896,
   "digest": "this section lists the sources used, including model cards, configuration files, research papers, and engine documentation, with dates of access provided.",
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-84ab2fdf",
   "q": "how do models define a token and a parameter?",
   "a": "A token is a word or a piece of one that a model reads and writes. A parameter, also called a weight, is one of the learned numbers that makes up a model.",
   "cites": [
    "the-words-this-page-leans-on"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-c1528ac4",
   "q": "how does a router function within a mixture of experts layer?",
   "a": "A router sits in front of the experts and scores each one for every token, picking the best few. The outputs of the picked experts are weighted by the router's scores and summed before the token moves to the next layer.",
   "cites": [
    "what-an-expert-is-and-who-picks"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-16496a30",
   "q": "what happens when a model backbone exceeds card capacity?",
   "a": "When a model does not fit, it spills. This can occur by keeping certain layers' experts in system memory or by keeping every expert in system memory and caching only the hottest ones on the card.",
   "cites": [
    "two-ways-to-spill"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "a-harness-that-rewrites-itself",
   "why": "shares ground with § Two numbers on every model card · § What the paper reports"
  },
  {
   "slug": "a-short-history-of-qwen",
   "why": "shares ground with § Two numbers on every model card · § The family under many of our numbers"
  },
  {
   "slug": "three-at-the-table",
   "why": "shares ground with § Two numbers on every model card · § What the label doesn't tell you"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThanks to Qwen (the model, its card and config, under the Qwen Community License 1.0); DeepSeek (the model, its card, config and papers, MIT); ISTA-DASLab (the GGUF builds of both models and their cards); Georgi Gerganov and the ggml authors (llama.cpp, MIT); the Strata engine's author (github.com/Niko1221/Strata, MIT); Mistral AI (Mixtral, its paper and its config); the authors of the papers quoted; Hugging Face (the cards, the file listings and the 2023 explainer); and NVIDIA (the published bandwidth). None of them owed us anything. No model was run and no card was touched for this page. A small (human) team asked the question and will approve this page; a fleet of AI agents read the sources, re-ran the arithmetic and drafted the prose under that team's rulings.",
 "kit": null,
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 1,
  "digests_written": 11,
  "digests_kept": 6,
  "chips_written": 8,
  "chips_kept": 3,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 0,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 1,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 11
  },
  "new_noun_tokens_checked": [
   "0.1.39",
   "0.921",
   "125B",
   "16B",
   "1991",
   "2017",
   "2021",
   "2023",
   "2024",
   "204.8",
   "4B",
   "51B",
   "552B",
   "6",
   "6000",
   "6B",
   "8B",
   "8x7B",
   "96",
   "Adaptive",
   "CPU",
   "DeepSeek",
   "DeepSeekMoE",
   "Experts",
   "Flash",
   "GB",
   "GB/s",
   "GiB",
   "KV",
   "Local",
   "MTP",
   "Mixtral",
   "Mixtures",
   "MoE",
   "NVIDIA",
   "PRO",
   "Qwen3.8-Flash-Next",
   "RAM",
   "RTX",
   "SSD",
   "Strata",
   "Switch",
   "Transformer",
   "V4.1"
  ],
  "new_noun_tokens_withheld": 1,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 221,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 51,
    "text_sha": "94f5df2eefbc",
    "figure": "0.921",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): 'the Strata engine 0.1.39 logged a hit rate of 0.921' cuts the sentence at its figure: the section gives 0.921 for the 24 GB desktop card alone (8,144 of 24,576 experts on the card), at a 4K prompt, as the share of expert lookups served from the card, and the line says neither which card, which prompt nor what the share is of; the same engine version ran on the 96 GB card with every expert on the card",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-04: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-04 (matched on its line key). The earlier record reads: removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): 'the Strata engine 0.1.39 logged a hit rate of 0.921' cuts the sentence at its figure: the section gives 0.921 for the 24 GB desktop card alone (8,144 of 24,576 experts on the card), at a 4K prompt, as the share of expert lookups served from the card, and the line …"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 230,
    "text_sha": "066958c7398a",
    "section": "why-the-big-number-decides-whether-it-fits",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): the digest says every parameter must be 'in RAM', where the section says every expert must be within reach 'on the card first of all' and quotes Hugging Face's RAM line as an outside source; this page separates the card's memory from system memory in its own word list, so the digest moves the section's point off the card",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-04: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-04 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): the digest says every parameter must be 'in RAM', where the section says every expert must be within reach 'on the card first of all' and quotes Hugging Face's RAM line as an outside source; this page separates the card's memory from system memory in its own word l…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 230,
    "text_sha": "4517898d31a0",
    "section": "two-ways-to-spill",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): two errors against the section. 'some layers are kept in the CPU': llama.cpp kept the EXPERTS of 34 of 48 layers in system memory and put the rest of every layer on the card. 'a router caches the hottest experts on the card': the caching is the Strata engine's, by its shipped profile ranked by routing frequency; the router is the model's own part that picks experts for each token",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-04: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-04 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): two errors against the section. 'some layers are kept in the CPU': llama.cpp kept the EXPERTS of 34 of 48 layers in system memory and put the rest of every layer on the card. 'a router caches the hottest experts on the card': the caching is the Strata engine's, by …"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 216,
    "text_sha": "88d7d98bd7c8",
    "section": "what-it-means-for-your-own-card",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): the digest states the table's first row (the whole model fits the card) as the rule, 'speed depends on the small number and card memory speed', and leaves out the three rows where it does not fit, where the processor's threads, system memory, the hit rate and the link set the speed; that difference is the section's point",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-04: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-04 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): the digest states the table's first row (the whole model fits the card) as the rule, 'speed depends on the small number and card memory speed', and leaves out the three rows where it does not fit, where the processor's threads, system memory, the hit rate and the l…"
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 262,
    "text_sha": "690cace659db",
    "chip_id": "c-4ff31c53",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): 'making it faster than a dense model' is stated as a rule; the section shows it for one pair on one card, cap and prompt (a 26B mixture against a dense 24B, 3.7 times as fast) and adds its own caution that the always-on parts roughly double the bytes a token touches",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-04: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-04 (matched on its question). The earlier record reads: removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): 'making it faster than a dense model' is stated as a rule; the section shows it for one pair on one card, cap and prompt (a 26B mixture against a dense 24B, 3.7 times as fast) and adds its own caution that the always-on parts roughly double the bytes a token touche…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 223,
    "text_sha": "c968ff2bcb33",
    "section": "why-the-small-number-decides-how-fast-it-writes-once-it-fits",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC), the lead's fold A2 of the release audit: 'writing speed is determined by how much data a card must move' states as a rule for every model what the section says of the trip, and leaves out what the section names as the setter for a dense model ('its writing speed is set by how fast the card can move its memory'); and 'mixture models can be faster than dense models' reads the one pair the section measured (a 26B mixture against a dense 24B on one NVIDIA GeForce RTX 3090, at one cap and with one prompt, 3.7 times as fast) as a property of the class, the reading this hand-read refused in chip c-4ff31c53",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC), the lead's fold A2 of the release audit: 'writing speed is determined by how much data a card must move' states as a rule for every model what the section says of the trip, and leaves out what the section names as the setter for a dense model ('its writing speed is set by how fast the card can move its memory'); and 'mixture models can be faster than dense models' reads the one pair the section measured (a 26B mixture against a dense 24B on one NVIDIA GeForce RTX 3090, at one cap and with one prompt, 3.7 times as fast) as a property of the class, the reading this hand-read refused in chip c-4ff31c53. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 320,
    "text_sha": "68ae42169122",
    "chip_id": "c-3642d981",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): 'activates 6 routed experts per token, resulting in 8B parameters during prefill and 16B during decode' adds a cause the section does not give: the 8B and 16B are DeepSeek's own two counts, whose difference the section says the paper's abstract explains only by naming a 'Causal Encoder-Decoder (CED) architecture', and this page goes no further",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): 'activates 6 routed experts per token, resulting in 8B parameters during prefill and 16B during decode' adds a cause the section does not give: the 8B and 16B are DeepSeek's own two counts, whose difference the section says the paper's abstract explains only by naming a 'Causal Encoder-Decoder (CED) architecture', and this page goes no further. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 354,
    "text_sha": "d413bf821a8a",
    "chip_id": "c-9fd66797",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): 'The 2021 Switch Transformer introduced sparsely-activated models' claims a first the section does not make: the section quotes the paper's description of itself, and its own 2017 line already puts a sparsely-gated layer of up to thousands of experts inside a language model",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): 'The 2021 Switch Transformer introduced sparsely-activated models' claims a first the section does not make: the section quotes the paper's description of itself, and its own 2017 line already puts a sparsely-gated layer of up to thousands of experts inside a language model. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 284,
    "text_sha": "2fe3546e2eff",
    "chip_id": "c-e313670b",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): the answer gives as the page's reason that 'all parameters need to be loaded in RAM', the error this hand-read dropped the section's digest for: the section says every expert must be within reach 'on the card first of all' and quotes Hugging Face's RAM line as an outside source, and the lookup tables, the exception the section names, live in system memory or on disk",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): the answer gives as the page's reason that 'all parameters need to be loaded in RAM', the error this hand-read dropped the section's digest for: the section says every expert must be within reach 'on the card first of all' and quotes Hugging Face's RAM line as an outside source, and the lookup tables, the exception the section names, live in system memory or on disk. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 242,
    "text_sha": "f53af62c3b67",
    "chip_id": "c-cf370068",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC): the answer states the table's first row (the whole model fits the card) as the rule for any card, the error this hand-read dropped the section's digest for: in the three rows where the model does not fit, the processor's threads, system memory, the hit rate and the link set the speed, and system memory must hold the spilled or the whole set of experts",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC): the answer states the table's first row (the whole model fits the card) as the rule for any card, the error this hand-read dropped the section's digest for: in the three rows where the model does not fit, the processor's threads, system memory, the hit rate and the link set the speed, and system memory must hold the spilled or the whole set of experts. Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 57,
    "text_sha": "fa68469607d0",
    "figure": "96",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-04",
    "why": "hand-read 2026-11-04 (UTC), the lead's ruling E6 on the audit of 164d2fe1: 'a card's label counts in powers of two so 96 GB is 96 GiB' is a definition, not a finding: it restates the page's word list (what a GiB is) rather than anything the page read or worked out, so it is no receipt line",
    "detail": "removed from this pack by hand on 2026-11-04 (UTC), after the writer ran and after the gates had counted: hand-read 2026-11-04 (UTC), the lead's ruling E6 on the audit of 164d2fe1: 'a card's label counts in powers of two so 96 GB is 96 GiB' is a definition, not a finding: it restates the page's word list (what a GiB is) rather than anything the page read or worked out, so it is no receipt line. Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-11-04T17:04:45Z"
}