{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "ram-speeds-compared",
 "title": "Speeds from RAM alone, compared",
 "dek": "The CPU-only writing, prompt-reading and first-word speeds we have measured since August, like with like, as bar charts: seven language models on four machines with no graphics card in the path, what more threads do, what a longer question costs, what the memory test read, and the 180-billion-parameter mixture that 128 GB of memory can hold. Every bar is read from a named file in a kit you can download.",
 "published": "2026-11-10",
 "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/ram-speeds-compared/",
  "md_url": "https://research.strata2signal.com/ram-speeds-compared/index.md",
  "md_sha": "bb1876d897acc6058bf819e8c66d091b1d31e02fc46c62ff80f2113a1bec85f1",
  "html_sha": "06b526c350bfea03e3795cfa4b005648c476573c5654721d4480605838b01aca",
  "anchors_sha": "6c6ebc70a80fff17505131c63abb286d49528522f023e5f3fe393afbe8eb8d7f"
 },
 "short": {
  "paragraph": "This page gathers every speed this house has measured for language models running from a computer's own memory, with no graphics card doing any of the work, as bar charts comparing machines, models, thread counts and question lengths. Four machines: a desktop with an AMD EPYC 7532 and 128 GB of eight-channel DDR4-3200, a laptop with an Intel Core Ultra 9 290HX Plus and 64 GB of DDR5-5600, a mini PC with an Intel Core i5-12600H and 32 GB of DDR5, and a desktop with an AMD Ryzen 7 3700X and 32 GB of DDR4; an older ten-core desktop with 48 GB of DDR4 stands in its own labelled figure. At the engine's default threads a 26-billion-parameter mixture of experts (gemma4:26b) wrote 12.66 to 13.64 tokens a second (a token is a little under a word) on the three machines with 32 or 64 GB and 24.68 on the EPYC; a 24-billion dense model (mistral-small3.2:24b) wrote 2.57 to 3.31 on the three and 7.09 on the EPYC. Ollama, the program that loads and runs the models, defaults to 4 threads on hybrid Intel chips: on the laptop, dense models wrote 1.75 to 1.91 times faster at 20, and both mixtures 36 to 44 per cent slower. The first word of a fresh answer to a 7,965-token question took 138.23 s on the Ryzen desktop and 146.22 s on the laptop (at 4 of its 24 threads, THROTTLED on every run), because the whole question is read first. Every figure is read from a file in a public kit; the labels a careful reader needs (THROTTLED, NOT WALL, default threads, fresh against warm) sit under each chart.",
  "from_dek": false,
  "counts": {
   "words": 9751,
   "minutes": 44,
   "tables": 21,
   "kit": true
  },
  "bullets": []
 },
 "sections": [
  {
   "id": "the-short-answer-two-models-on-four-machines",
   "heading": "The short answer: two models on four machines",
   "level": 2,
   "span": [
    2187,
    10091
   ],
   "chunks": [
    [
     2187,
     10091
    ]
   ],
   "chars": 7904,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-fast-does-each-model-write-on-each-machine",
   "heading": "How fast does each model write on each machine?",
   "level": 2,
   "span": [
    10091,
    13831
   ],
   "chunks": [
    [
     10091,
     13831
    ]
   ],
   "chars": 3740,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "should-i-raise-the-thread-count",
   "heading": "Should I raise the thread count?",
   "level": 2,
   "span": [
    13831,
    16475
   ],
   "chunks": [
    [
     13831,
     16475
    ]
   ],
   "chars": 2644,
   "digest": "writing speed is measured at default and alternative thread settings for a laptop, a mini PC, and a Ryzen 7 3700X. the results show how changing the thread count affects the tokens a second while writing for various models.",
   "digest_skipped": null
  },
  {
   "id": "how-long-until-the-first-word-of-the-answer-time-to-first",
   "heading": "How long until the first word of the answer? (time to first token, TTFT)",
   "level": 2,
   "span": [
    16475,
    20305
   ],
   "chunks": [
    [
     16475,
     20305
    ]
   ],
   "chars": 3830,
   "digest": "this section measures seconds from sending a fresh question to the first word of the answer for several models. the model is already loaded and nothing is cached, using default threads across four machines.",
   "digest_skipped": null
  },
  {
   "id": "how-fast-does-each-machine-read-the-question-prompt",
   "heading": "How fast does each machine read the question? (prompt processing)",
   "level": 2,
   "span": [
    20305,
    24123
   ],
   "chunks": [
    [
     20305,
     24123
    ]
   ],
   "chars": 3818,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "why-do-some-pages-claim-a-0-2-second-first-word",
   "heading": "Why do some pages claim a 0.2-second first word?",
   "level": 2,
   "span": [
    24123,
    25743
   ],
   "chunks": [
    [
     24123,
     25743
    ]
   ],
   "chars": 1620,
   "digest": "the section explains that Ollama keeps a prompt's computed prefix, meaning a resent prompt reports a near-zero reading time. this is a cache artifact where the first word can be reported much sooner than a fresh prompt.",
   "digest_skipped": null
  },
  {
   "id": "do-more-threads-shorten-the-wait",
   "heading": "Do more threads shorten the wait?",
   "level": 2,
   "span": [
    25743,
    35026
   ],
   "chunks": [
    [
     25743,
     35026
    ]
   ],
   "chars": 9283,
   "digest": "prompt reading and first word times are measured at two thread settings. the results for the laptop and Ryzen 7 3700X show how different thread counts impact the wait for the first word and the speed of prompt reading.",
   "digest_skipped": null
  },
  {
   "id": "is-a-mixture-of-experts-moe-really-faster-from-ram",
   "heading": "Is a mixture of experts (MoE) really faster from RAM?",
   "level": 2,
   "span": [
    35026,
    37555
   ],
   "chunks": [
    [
     35026,
     37555
    ]
   ],
   "chars": 2529,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "how-long-is-the-wait-for-a-long-question",
   "heading": "How long is the wait for a long question?",
   "level": 2,
   "span": [
    37555,
    39978
   ],
   "chunks": [
    [
     37555,
     39978
    ]
   ],
   "chars": 2423,
   "digest": "the wait for the first word is measured as the prompt grows from 0.5K to 8K tokens. the test uses two machines at default threads with fresh loads for gemma4:26b and qwen3.5:9b.",
   "digest_skipped": null
  },
  {
   "id": "does-a-longer-question-slow-the-writing",
   "heading": "Does a longer question slow the writing?",
   "level": 2,
   "span": [
    39978,
    41902
   ],
   "chunks": [
    [
     39978,
     41902
    ]
   ],
   "chars": 1924,
   "digest": "writing speed is measured after prompts of four different lengths. the test tracks tokens a second for gemma4:26b and qwen3.5:9b on the 290HX Plus laptop and the 3700X desktop.",
   "digest_skipped": null
  },
  {
   "id": "does-a-longer-question-slow-the-reading",
   "heading": "Does a longer question slow the reading?",
   "level": 2,
   "span": [
    41902,
    46361
   ],
   "chunks": [
    [
     41902,
     46361
    ]
   ],
   "chars": 4459,
   "digest": "prompt reading speed is measured at four different question lengths. the results for gemma4:26b and qwen3.5:9b show how the reading rate changes as the prompt size increases on the laptop and desktop.",
   "digest_skipped": null
  },
  {
   "id": "where-do-more-threads-stop-helping",
   "heading": "Where do more threads stop helping?",
   "level": 2,
   "span": [
    46361,
    48758
   ],
   "chunks": [
    [
     46361,
     48758
    ]
   ],
   "chars": 2397,
   "digest": "writing speed is measured at various thread counts for Nemotron 30B-A3B on a mini PC and qwen3.5:9b on a laptop and mini PC. the results show how increasing or decreasing threads affects tokens a second.",
   "digest_skipped": null
  },
  {
   "id": "what-did-the-stream-triad-read-on-each-machine",
   "heading": "What did the STREAM triad read on each machine?",
   "level": 2,
   "span": [
    48758,
    50882
   ],
   "chunks": [
    [
     48758,
     50882
    ]
   ],
   "chars": 2124,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "does-writing-speed-follow-the-triad",
   "heading": "Does writing speed follow the triad?",
   "level": 2,
   "span": [
    50882,
    57412
   ],
   "chunks": [
    [
     50882,
     57412
    ]
   ],
   "chars": 6530,
   "digest": "the section compares what the triad read to what gemma4:26b wrote on the same machine and day. it uses a bandwidth rule k to see if the writing speed follows the measured memory speed.",
   "digest_skipped": null
  },
  {
   "id": "what-does-128-gb-of-eight-channel-memory-buy",
   "heading": "What does 128 GB of eight-channel memory buy?",
   "level": 2,
   "span": [
    57412,
    59943
   ],
   "chunks": [
    [
     57412,
     59943
    ]
   ],
   "chars": 2531,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-does-the-laptops-processor-draw-while-writing-the",
   "heading": "What does the laptop's processor draw while writing? (the frozen prompt, 2026-11-05)",
   "level": 2,
   "span": [
    59943,
    61342
   ],
   "chunks": [
    [
     59943,
     61342
    ]
   ],
   "chars": 1399,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "and-on-the-long-questions-the-prompt-size-ladder-2026-11-08",
   "heading": "And on the long questions? (the prompt-size ladder, 2026-11-08)",
   "level": 2,
   "span": [
    61342,
    62763
   ],
   "chunks": [
    [
     61342,
     62763
    ]
   ],
   "chars": 1421,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "and-the-epycs-on-the-big-model",
   "heading": "And the EPYC's, on the big model?",
   "level": 2,
   "span": [
    62763,
    65326
   ],
   "chunks": [
    [
     62763,
     65326
    ]
   ],
   "chars": 2563,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "an-older-desktop-ten-cores-48-gb-of-ddr4",
   "heading": "An older desktop: ten cores, 48 GB of DDR4",
   "level": 2,
   "span": [
    65326,
    67721
   ],
   "chunks": [
    [
     65326,
     67721
    ]
   ],
   "chars": 2395,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "embeddings-on-the-processor-for-search-and-rag",
   "heading": "Embeddings on the processor (for search and RAG)",
   "level": 2,
   "span": [
    67721,
    69199
   ],
   "chunks": [
    [
     67721,
     69199
    ]
   ],
   "chars": 1478,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-the-numbers-do-not-say",
   "heading": "What the numbers do not say",
   "level": 2,
   "span": [
    69199,
    70674
   ],
   "chunks": [
    [
     69199,
     70674
    ]
   ],
   "chars": 1475,
   "digest": "this section lists several limitations of the data, including the lack of quality readings, the absence of a controlled DDR4-against-DDR5 pair, and the fact that no wall power was measured.",
   "digest_skipped": null
  },
  {
   "id": "not-on-this-page",
   "heading": "Not on this page",
   "level": 2,
   "span": [
    70674,
    71356
   ],
   "chunks": [
    [
     70674,
     71356
    ]
   ],
   "chars": 682,
   "digest": "this section identifies tasks and data that are not included on this page, such as other processor timings and the mini PC's wall reading.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    71356,
    72260
   ],
   "chunks": [
    [
     71356,
     72260
    ]
   ],
   "chars": 904,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    72260,
    72793
   ],
   "chunks": [
    [
     72260,
     72793
    ]
   ],
   "chars": 533,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    72793,
    73400
   ],
   "chunks": [
    [
     72793,
     73400
    ]
   ],
   "chars": 607,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "sources",
   "heading": "Sources",
   "level": 2,
   "span": [
    73400,
    73925
   ],
   "chunks": [
    [
     73400,
     73925
    ]
   ],
   "chars": 525,
   "digest": "the sources for the data include the three kits and specific articles from the house.",
   "digest_skipped": null
  }
 ],
 "chips": [
  {
   "id": "c-993d6dec",
   "q": "how does changing thread counts affect writing speed?",
   "a": "Increasing threads can improve writing speed for some models, such as qwen3.5:9b on a laptop, which moves from 7.45 tok/s at default to 13.05 tok/s at 20 threads. However, for gemma4:26b on a laptop, increasing from 4 to 20 threads actually reduces writing speed from 13.26 tok/s to 8.44 tok/s.",
   "cites": [
    "should-i-raise-the-thread-count"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-3e3ed0ce",
   "q": "how does thread count influence writing speed in sweeps?",
   "a": "In some cases, more threads decrease speed. For the nemotron-3.5-lightning:30b-a3b model on a mini PC, writing speed drops from 14.44 tok/s at 4 threads to 1.27 tok/s at 16 threads.",
   "cites": [
    "where-do-more-threads-stop-helping"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "six-cards-the-same-questions",
   "why": "shares ground with § Does a longer question slow the writing? · § Writing: how fast each one answers"
  },
  {
   "slug": "flash-next-on-three-cards",
   "why": "shares ground with § What the numbers do not say · § What it does not say"
  },
  {
   "slug": "two-hours-on-battery",
   "why": "shares ground with § Where do more threads stop helping? · § Set the thread count — the one free lever, and the exception that proves the rule"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nThe models ran on Ollama 0.32.15 (0.32.13 on the older desktop) with llama.cpp beneath it, the big mixture on llama.cpp build 11514. They are Alibaba's Qwen3.5 and Qwen3.8-Flash-Next, Google's Gemma 4, Microsoft's Phi-4, Mistral Small 3.2 from Mistral AI, NVIDIA's Nemotron 3.5 Lightning, Meta's Llama 3.3 and Nomic's nomic-embed-text; each earlier page names the licence it found. A small human team pasted the stages, read the memory slots and signed the amendments; AI agents built the kits, drew the charts and wrote this page under its rulings.",
 "kit": {
  "url": "https://research.strata2signal.com/ram-speeds-compared/data/",
  "licence": "CC BY 4.0",
  "files": [
   "README.md",
   "charts/an-older-desktop.json",
   "charts/embeddings-on-the-processor.json",
   "charts/epyc-package-watts-big-model.json",
   "charts/first-word-matrix.json",
   "charts/laptop-package-watts-frozen-prompt.json",
   "charts/laptop-package-watts-ladder.json",
   "charts/mixture-against-dense.json",
   "charts/prompt-reading-by-question-length.json",
   "charts/prompt-reading-matrix.json",
   "charts/the-big-model-from-128-gb.json",
   "charts/the-cached-prompt-trap.json",
   "charts/the-short-answer.json",
   "charts/the-wait-for-a-long-question.json",
   "charts/threads-and-the-wait.json",
   "charts/threads-and-writing-speed.json",
   "charts/what-the-triad-read.json",
   "charts/where-more-threads-stop-helping.json",
   "charts/writing-speed-and-the-triad.json",
   "charts/writing-speed-by-question-length.json",
   "charts/writing-speed-matrix.json",
   "facts/FACTS.md",
   "fill-in/laptop/RECEIPT.txt",
   "fill-in/laptop/rows.jsonl",
   "fill-in/laptop/summary.json",
   "fill-in/laptop/triad.json",
   "fill-in/memory-reads.md",
   "fill-in/prompts/PROMPTS.json",
   "fill-in/prompts/fillin-p0500.txt",
   "fill-in/prompts/fillin-p2000.txt",
   "fill-in/prompts/fillin-p4000.txt",
   "fill-in/prompts/fillin-p8000.txt",
   "fill-in/two-channel-desktop/RECEIPT.txt",
   "fill-in/two-channel-desktop/rows.jsonl",
   "fill-in/two-channel-desktop/summary.json",
   "fill-in/two-channel-desktop/triad.json",
   "older-desktop/gpu-3080-m1-cpu-box.json",
   "older-desktop/gpu-3080-m2-cpu-box.json",
   "older-desktop/gpu-3080-m3-cpu-box.json",
   "older-desktop/gpu-3080-m4-cpu-box.json",
   "older-desktop/gpu-3080-m5-cpu-box.json",
   "older-desktop/gpu-3080-m6-cpu-box.json",
   "older-desktop/gpu-3080-x1-cpu-box.json",
   "older-desktop/gpu-3080-x2-cpu-box.json",
   "older-desktop/gpu-3080-x3-cpu-box.json",
   "older-desktop/gpu-3080-x4-cpu-box.json"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 0,
  "digests_written": 23,
  "digests_kept": 12,
  "chips_written": 8,
  "chips_kept": 2,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 1,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 9,
   "profanity": 0,
   "identifier": 0,
   "bare_figure": 0,
   "house_voice": 0,
   "editorial": 0,
   "hand-edit": 10
  },
  "new_noun_tokens_checked": [
   "0.5K",
   "1.27",
   "100",
   "100.69",
   "12.12",
   "128",
   "12b",
   "13.05",
   "13.26",
   "13.5",
   "13.64",
   "14.44",
   "14.94",
   "146.22",
   "14b",
   "15.87",
   "16",
   "180B",
   "20",
   "24.63",
   "24.68",
   "24b",
   "26.50",
   "26b",
   "290HX",
   "3.74",
   "30B-A3B",
   "30b-a3b",
   "3700X",
   "4",
   "54.5",
   "64.3",
   "7",
   "7.09",
   "7.45",
   "7.72",
   "70b",
   "7532",
   "8.16",
   "8.44",
   "8K",
   "9b",
   "DDR4-against-DDR5",
   "GB",
   "GB/s",
   "Nemotron",
   "Ollama",
   "PC",
   "PC's",
   "Plus",
   "Q4KM",
   "Q80",
   "Qwen3.8-Flash-Next",
   "RAPL",
   "Ryzen",
   "STREAM",
   "W",
   "gemma4",
   "i5-12600H",
   "llama3.3",
   "mistral-small3.2",
   "nemotron-3.5-lightning",
   "package-0",
   "phi4",
   "qwen3.5"
  ],
  "new_noun_tokens_withheld": 9,
  "figure_definition": "v3",
  "figure_boundary": "guarded",
  "drops": [
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 255,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "digest",
    "reason": "figure",
    "item": "mixture of experts models are compared to dense models on the same machine at default threads and at 20 threads. the comparison includes gemma4:26b against mistral-small3.2:24b and Nemotron 30B-A3B ag",
    "detail": "figures absent: ['3.2:24', '3.5:9']"
   },
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 170,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 207,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 178,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 162,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "bullet",
    "reason": "redaction",
    "item": null,
    "item_chars": 58,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "chip",
    "reason": "redaction",
    "item": null,
    "item_chars": 315,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "chip",
    "reason": "redaction",
    "item": null,
    "item_chars": 250,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "chip",
    "reason": "redaction",
    "item": null,
    "item_chars": 236,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 216,
    "text_sha": "ce9f61318262",
    "section": "how-fast-does-each-model-write-on-each-machine",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "says the section's data 'includes different prompt sizes'; this section is the frozen-prompt bench, one 2,101-byte prompt, and the prompt-size ladder is another family the page keeps apart (hand-read 2026-11-10, UTC)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-10: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-10 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: says the section's data 'includes different prompt sizes'; this section is the frozen-prompt bench, one 2,101-byte prompt, and the prompt-size ladder is another family the page keeps apart (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 173,
    "text_sha": "fbbb13699b0d",
    "section": "how-fast-does-each-machine-read-the-question-prompt",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "says the section 'covers different prompt sizes'; it is one prompt, the same 2,101 bytes read as 505 to 1,023 tokens by each model's tokenizer, which is why the page says the panels are not comparable per token (hand-read 2026-11-10, UTC)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-10: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-10 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: says the section 'covers different prompt sizes'; it is one prompt, the same 2,101 bytes read as 505 to 1,023 tokens by each model's tokenizer, which is why the page says the panels are not comparable per token (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds…"
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 164,
    "text_sha": "ba41f70d0f4e",
    "section": "what-does-the-laptops-processor-draw-while-writing-the",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "speaks of the processor's package watts without NOT WALL, the label the page prints on every package-watt figure, and writes the THROTTLED label as a verb (hand-read 2026-11-10, UTC)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-10: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-10 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: speaks of the processor's package watts without NOT WALL, the label the page prints on every package-watt figure, and writes the THROTTLED label as a verb (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 170,
    "text_sha": "5c78a6216c7f",
    "section": "and-on-the-long-questions-the-prompt-size-ladder-2026-11-08",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "speaks of the processor's package watts without NOT WALL, the label the page prints on every package-watt figure (hand-read 2026-11-10, UTC)",
    "detail": "Carried forward by --carry-hand-edits on 2026-11-10: the writer produced this line again and the pack on disk records it removed by hand on 2026-11-10 (matched on its section). The earlier record reads: removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: speaks of the processor's package watts without NOT WALL, the label the page prints on every package-watt figure (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 42,
    "text_sha": "8b347b1a1b61",
    "figure": "13.5",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "a package-watt figure without NOT WALL, the label the page prints on every package-watt figure, and without its condition: the 13.5 W is the laptop's package at idle for 60 s, not while writing, as the line reads (hand-read 2026-11-10, UTC)",
    "detail": "removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: a package-watt figure without NOT WALL, the label the page prints on every package-watt figure, and without its condition: the 13.5 W is the laptop's package at idle for 60 s, not while writing, as the line reads (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "bullet",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 64,
    "text_sha": "56c2468934ff",
    "figure": "13.64",
    "withheld": "one bullet removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "a bar read as a finding without its bench marker: the Ryzen desktop's 13.64 tok/s is a fresh-load November run (n = 3) at the engine's default 8 threads, which the writing matrix sets beside 2026-08-25 warm runs, and the page says a gap under about 16 % between those benches is not a finding; the line drops both conditions (hand-read 2026-11-10, UTC)",
    "detail": "removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: a bar read as a finding without its bench marker: the Ryzen desktop's 13.64 tok/s is a fresh-load November run (n = 3) at the engine's default 8 threads, which the writing matrix sets beside 2026-08-25 warm runs, and the page says a gap under about 16 % between those benches is not a finding; the line drops both conditions (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "digest",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 205,
    "text_sha": "0a03cf548267",
    "section": "an-older-desktop-ten-cores-48-gb-of-ddr4",
    "withheld": "one section digest removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "lists gemma4:26b among the older desktop's results without the section's condition that it is an older build of the model, not the build of every other bar, and drops the older engine build and 'its own figure, joined to nothing above', the conditions the section exists to state (hand-read 2026-11-10, UTC)",
    "detail": "removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: lists gemma4:26b among the older desktop's results without the section's condition that it is an older build of the model, not the build of every other bar, and drops the older engine build and 'its own figure, joined to nothing above', the conditions the section exists to state (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 236,
    "text_sha": "509a01ddf662",
    "chip_id": "c-f2303719",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "prints the laptop's ladder first words (7.72 s, 146.22 s) without THROTTLED, the label the page prints beside them: every one of those gemma4:26b runs is THROTTLED 3/3 (T4a); the earlier hand-read's c-9d2f7daf claim, back under a new id (hand-read 2026-11-10, UTC)",
    "detail": "removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: prints the laptop's ladder first words (7.72 s, 146.22 s) without THROTTLED, the label the page prints beside them: every one of those gemma4:26b runs is THROTTLED 3/3 (T4a); the earlier hand-read's c-9d2f7daf claim, back under a new id (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 219,
    "text_sha": "131fc9ef7052",
    "chip_id": "c-39a3a5de",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "prints the laptop's ladder writing speeds (14.94, 8.16 tok/s) without THROTTLED (3/3 on every run, T4a), widens the fall to every model, and drops the section's condition that the mixture's steeper fall is not a property of mixtures (its draft acceptance moves across the texts); the earlier hand-read's c-1ed88618 claim, back under a new id (hand-read 2026-11-10, UTC)",
    "detail": "removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: prints the laptop's ladder writing speeds (14.94, 8.16 tok/s) without THROTTLED (3/3 on every run, T4a), widens the fall to every model, and drops the section's condition that the mixture's steeper fall is not a property of mixtures (its draft acceptance moves across the texts); the earlier hand-read's c-1ed88618 claim, back under a new id (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   },
   {
    "kind": "chip",
    "reason": "hand-edit",
    "item": null,
    "item_chars": 212,
    "text_sha": "b858fa80ddc1",
    "chip_id": "c-af466d27",
    "withheld": "one chip removed from this pack by hand",
    "removed_utc": "2026-11-10",
    "why": "prints the laptop's ladder reading rates (64.3, 54.5 tok/s) without THROTTLED (3/3 on every run, T4a) and drops the section's condition that the laptop's decline cannot be separated from heat; the earlier hand-read's c-8a86feb2 claim, back under a new id (hand-read 2026-11-10, UTC)",
    "detail": "removed from this pack by hand on 2026-11-10 (UTC), after the writer ran and after the gates had counted: prints the laptop's ladder reading rates (64.3, 54.5 tok/s) without THROTTLED (3/3 on every run, T4a) and drops the section's condition that the laptop's decline cannot be separated from heat; the earlier hand-read's c-8a86feb2 claim, back under a new id (hand-read 2026-11-10, UTC). Not a gate drop, and recorded here so the ledger adds up."
   }
  ]
 },
 "generated_utc": "2026-11-10T07:05:10Z"
}