{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "the-instrument-travels",
 "title": "The Instrument Travels",
 "dek": "Carry a frozen exam from a 96 GB workstation card to a 24 GB consumer one and the model it ranked two weeks ago ranks again — at the floor, on a different build, with nothing new seated.",
 "published": "2026-08-29",
 "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-instrument-travels/",
  "md_url": "https://research.strata2signal.com/the-instrument-travels/index.md",
  "md_sha": "530098f413b700842fa3392f5a9021d0a1123aeda7b3f540257e43448b60d99e",
  "html_sha": "a2556e1b36707846b3e336030cf2eedacefac426821ee10d5d0243ed668bff71",
  "anchors_sha": "b65d3a2be57d6a46667eafa00bb463253737c1f0a806360edbdfec2aad68cb96"
 },
 "short": {
  "paragraph": "Carry a frozen exam from a 96 GB workstation card to a 24 GB consumer one and the model it ranked two weeks ago ranks again - at the floor, on a different build, with nothing new seated.",
  "from_dek": true,
  "counts": {
   "words": 4490,
   "minutes": 20,
   "tables": 5,
   "kit": true
  },
  "bullets": [
   {
    "text": "the battery stopped itself at 09:58Z",
    "figure": "09:58Z",
    "cite": "what-we-ran-and-the-rule-behind-every-number",
    "quote": "The battery stopped itself at 09:58Z, mid-leg, when free space on the bench box's disk fell below the harness's 60 GiB safety floor.",
    "span": [
     5283,
     5290
    ]
   },
   {
    "text": "the judge seat's trial runs at a context of 16,384-token",
    "figure": "16,384-token",
    "cite": "the-words-these-tables-use-and-the-floors-they-score-against",
    "quote": "The judge seat's 43-case trial is an older, separately frozen exam and runs at its own registered settings - a 16,384-token context and no think flag at all - where every other instrument here runs 32,768 with think:false explicit.",
    "span": [
     10352,
     10365
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "the-chair-and-who-gets-to-sit-in-it",
   "heading": "The chair, and who gets to sit in it",
   "level": 2,
   "span": [
    1768,
    3063
   ],
   "chunks": [
    [
     1768,
     3063
    ]
   ],
   "chars": 1295,
   "digest": "the article introduces the concept of a chair as a model position earned through exams. it discusses the necessity of testing newcomers on the same hardware and exams as incumbents to ensure fair comparison and to see if the instrument travels.",
   "digest_skipped": null
  },
  {
   "id": "what-we-ran-and-the-rule-behind-every-number",
   "heading": "What we ran, and the rule behind every number",
   "level": 2,
   "span": [
    3063,
    6627
   ],
   "chunks": [
    [
     3063,
     6627
    ]
   ],
   "chars": 3564,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-words-these-tables-use-and-the-floors-they-score-against",
   "heading": "The words these tables use, and the floors they score against",
   "level": 2,
   "span": [
    6627,
    10954
   ],
   "chunks": [
    [
     6627,
     10954
    ]
   ],
   "chars": 4327,
   "digest": "this section defines terms like seat, posture, and various exam types. it explains scoring metrics such as kill-recall and preservation, and details the pre-registered floors and the specific settings used for the judge seat's trial.",
   "digest_skipped": null
  },
  {
   "id": "the-standings",
   "heading": "The standings",
   "level": 2,
   "span": [
    10954,
    15440
   ],
   "chunks": [
    [
     10954,
     15440
    ]
   ],
   "chars": 4486,
   "digest": "tables present the results of the schema probe, judge trial, judge seat exam, and other tests. results include counts for kill-recall, preservation, and speed, while noting which models were seat-blocked, unmeasurable, or not-run.",
   "digest_skipped": null
  },
  {
   "id": "what-the-battery-was-for-the-instrument-travels",
   "heading": "What the battery was for: the instrument travels",
   "level": 2,
   "span": [
    15440,
    18330
   ],
   "chunks": [
    [
     15440,
     18330
    ]
   ],
   "chars": 2890,
   "digest": "the section examines how results for qwen3.6:27b compare to previous runs. it notes that while some counts like kill-recall remained identical, the specific missed item moved, and discusses how different instruments reproduced results differently.",
   "digest_skipped": null
  },
  {
   "id": "schema-perfect-and-still-blocked",
   "heading": "Schema-perfect, and still blocked",
   "level": 2,
   "span": [
    18330,
    19354
   ],
   "chunks": [
    [
     18330,
     19354
    ]
   ],
   "chars": 1024,
   "digest": "muse-glimmer:30b is described as schema-perfect but seat-blocked because a trailing end-of-text sentinel in the bytes breaks a strict parser, despite the model passing a long-context filing exam.",
   "digest_skipped": null
  },
  {
   "id": "the-judges-chair-scored-one-candidate-of-three",
   "heading": "The judge's chair scored one candidate of three",
   "level": 2,
   "span": [
    19354,
    21732
   ],
   "chunks": [
    [
     19354,
     21732
    ]
   ],
   "chars": 2378,
   "digest": "of three candidates that sat the judge seat's 43-case trial, only gemma4:31b was scored. gemma4:31b cleared the kill floor but failed preservation. other models were unmeasurable due to truncations against the 1,024-token reply budget.",
   "digest_skipped": null
  },
  {
   "id": "what-a-24-gb-door-actually-filters",
   "heading": "What a 24 GB door actually filters",
   "level": 2,
   "span": [
    21732,
    22966
   ],
   "chunks": [
    [
     21732,
     22966
    ]
   ],
   "chars": 1234,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "the-row-that-owes-a-re-run",
   "heading": "The row that owes a re-run",
   "level": 2,
   "span": [
    22966,
    23870
   ],
   "chunks": [
    [
     22966,
     23870
    ]
   ],
   "chars": 904,
   "digest": "laguna-xs-2.1:latest showed an inverted trap where schema output is clean under think:false but drops the constraint under think:true. the model's trial was cut mid-leg due to the disk stop and time-box.",
   "digest_skipped": null
  },
  {
   "id": "what-this-page-does-not-say",
   "heading": "What this page does not say",
   "level": 2,
   "span": [
    23870,
    25585
   ],
   "chunks": [
    [
     23870,
     25585
    ]
   ],
   "chars": 1715,
   "digest": "the article clarifies that no seats were awarded, explains a speed tier anomaly for glimmer, notes that speed measurements were taken on a shared host, and reiterates that weights are re-pullable but measurements are not.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    25585,
    26870
   ],
   "chunks": [
    [
     25585,
     26870
    ]
   ],
   "chars": 1285,
   "digest": "the summary emphasizes that counts travel while rates do not, a 24 GB door filters by artifact, and the bench stopped itself to protect the disk. it also notes that frozen means frozen, including inherited exams.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work",
   "heading": "How to check our work",
   "level": 2,
   "span": [
    26870,
    28250
   ],
   "chunks": [
    [
     26870,
     28250
    ]
   ],
   "chars": 1380,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    28250,
    29343
   ],
   "chunks": [
    [
     28250,
     29343
    ]
   ],
   "chars": 1093,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-9f5df00c",
   "q": "how are model rankings determined in this context?",
   "a": "A model earns a chair by passing the same exams as the incumbent on hardware a normal person owns. The process compares whether a newcomer can pass the same tests as previous models to ensure the results are consistent and the instrument travels.",
   "cites": [
    "the-chair-and-who-gets-to-sit-in-it"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-5c867995",
   "q": "what rules govern the comparison of model counts and rates?",
   "a": "A count, such as kill-recall or task success, is a property of the model and travels across different hardware. A rate, such as tokens per second, is a property of both the model and the silicon and does not travel.",
   "cites": [
    "what-we-ran-and-the-rule-behind-every-number"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-aac30d01",
   "q": "what criteria define a successful judge trial?",
   "a": "A judge trial passes if the model achieves a kill-recall of at least 11 of 12 and a preservation of at least 8 of 9. For the judge seat's trial, the requirements are a kill-recall of at least 23 of 27 and a preservation of at least 13 of 16.",
   "cites": [
    "the-words-these-tables-use-and-the-floors-they-score-against"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-8923d06d",
   "q": "which candidate achieved a ranked verdict in the judge trial?",
   "a": "qwen3.6:27b achieved a RANKED verdict in the judge trial with a kill-recall of 11 and a preservation of 8. Other candidates were either exploratory, unmeasurable, or did not complete the run due to a time-box or disk halt.",
   "cites": [
    "the-standings"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-8264b215",
   "q": "how does a model's build affect its reproducibility?",
   "a": "Reproducibility is demonstrated when a different build of a model achieves similar results on different silicon. For qwen3.6:27b, the kill-recall remained 11/12 despite a different manifest digest, showing the instrument travels even when the model moves slightly.",
   "cites": [
    "what-the-battery-was-for-the-instrument-travels"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-547cf5e3",
   "q": "why might a model be labeled as seat-blocked despite perfect schema output?",
   "a": "A model is seat-blocked if it emits schema-perfect JSON but the runtime leaks a trailing end-of-text sentinel into the bytes, which breaks a strict parser. This is a failure of the contract rather than a failure of the model's architecture.",
   "cites": [
    "schema-perfect-and-still-blocked"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "the-new-kid",
   "why": "shares ground with § The judge's chair scored one candidate of three · § The judge seat — and the floor it missed"
  },
  {
   "slug": "outside-judges",
   "why": "shares ground with § The standings · § The gauntlet: our judges, examined"
  },
  {
   "slug": "seat-trials",
   "why": "shares ground with § The chair, and who gets to sit in it · § Where the chair sits"
  }
 ],
 "thanks": "",
 "kit": {
  "url": "https://research.strata2signal.com/the-instrument-travels/data/",
  "licence": "CC BY 4.0",
  "files": [
   "README.md",
   "TEARDOWN-RECEIPT.md",
   "battery.log",
   "c2-scores.json",
   "c3-scores.json",
   "c5-recovered.json",
   "c7-muse-glimmer_30b.json",
   "c7-qwen3.6_27b.json",
   "counting-rules.md",
   "digests.md",
   "field/gemma4-31b.json",
   "field/muse-glimmer-30b.json",
   "field/qwen3.6-27b.json",
   "fit-gemma4_31b.json",
   "fit-laguna-xs-2.1_latest.json",
   "fit-muse-glimmer_30b.json",
   "fit-nemotron-3.5-lightning_30b.json",
   "fit-qwen3.6_27b.json",
   "prereg/GOLDEN-SHAS.md",
   "prereg/PREREG-C1.md",
   "probes-gemma4_31b.json",
   "probes-laguna-xs-2.1_latest.json",
   "probes-muse-glimmer_30b.json",
   "probes-qwen3.6_27b.json",
   "provenance.json",
   "raw/C1/gemma4_31b/C1-gemma4_31b.jsonl",
   "raw/C1/laguna-xs-2.1_latest/C1-laguna-xs-2.1_latest.jsonl",
   "raw/C1/muse-glimmer_30b/C1-muse-glimmer_30b.jsonl",
   "raw/C1/qwen3.6_27b/C1-qwen3.6_27b.jsonl",
   "raw/C5/dflash-pair/C5-dflash-pair.jsonl",
   "raw/C5/muse-glimmer_30b/C5-muse-glimmer_30b.jsonl",
   "raw/C5/qwen3.6_27b/C5-qwen3.6_27b.jsonl",
   "raw/c2/gemma4_31b/c2-ctx-24k--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-ctx-24k--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-ctx-8k--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-ctx-8k--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-cw-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-cw-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-cw-2--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-cw-2--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-cw-3--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-cw-3--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-cw-4--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-cw-4--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-hon-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-hon-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-hon-2--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-hon-2--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-hon-3--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-hon-3--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-hon-4--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-hon-4--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-json-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-json-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-json-2--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-json-2--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-num-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-num-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-num-2--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-num-2--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-num-3--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-num-3--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-re-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-re-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-sh-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-sh-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-sum-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-sum-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-sum-2--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-sum-2--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/c2-tone-1--thinkfalse--r1.json",
   "raw/c2/gemma4_31b/c2-tone-1--thinkfalse--r2.json",
   "raw/c2/gemma4_31b/calls.jsonl",
   "raw/c2/muse-glimmer_30b/c2-ctx-24k--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-ctx-24k--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-ctx-8k--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-ctx-8k--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-cw-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-cw-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-cw-2--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-cw-2--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-cw-3--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-cw-3--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-cw-4--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-cw-4--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-hon-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-hon-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-hon-2--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-hon-2--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-hon-3--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-hon-3--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-hon-4--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-hon-4--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-json-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-json-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-json-2--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-json-2--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-num-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-num-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-num-2--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-num-2--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-num-3--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-num-3--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-re-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-re-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-sh-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-sh-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-sum-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-sum-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-sum-2--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-sum-2--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/c2-tone-1--thinkfalse--r1.json",
   "raw/c2/muse-glimmer_30b/c2-tone-1--thinkfalse--r2.json",
   "raw/c2/muse-glimmer_30b/calls.jsonl",
   "raw/c2/qwen3.6_27b/c2-ctx-24k--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-ctx-24k--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-ctx-8k--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-ctx-8k--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-cw-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-cw-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-cw-2--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-cw-2--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-cw-3--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-cw-3--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-cw-4--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-cw-4--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-hon-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-hon-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-hon-2--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-hon-2--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-hon-3--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-hon-3--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-hon-4--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-hon-4--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-json-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-json-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-json-2--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-json-2--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-num-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-num-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-num-2--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-num-2--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-num-3--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-num-3--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-re-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-re-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-sh-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-sh-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-sum-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-sum-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-sum-2--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-sum-2--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/c2-tone-1--thinkfalse--r1.json",
   "raw/c2/qwen3.6_27b/c2-tone-1--thinkfalse--r2.json",
   "raw/c2/qwen3.6_27b/calls.jsonl",
   "raw/c2/run-manifest.json",
   "raw/c3/gemma4_31b/preprobe.json",
   "raw/c3/gemma4_31b/think_false/calls.jsonl",
   "raw/c3/gemma4_31b/think_false/trace.jsonl",
   "raw/c3/manifest.json",
   "raw/c3/muse-glimmer_30b/preprobe.json",
   "raw/c3/muse-glimmer_30b/think_false/calls.jsonl",
   "raw/c3/muse-glimmer_30b/think_false/trace.jsonl",
   "raw/c3/qwen3.6_27b/preprobe.json",
   "raw/c3/qwen3.6_27b/think_false/calls.jsonl",
   "raw/c3/qwen3.6_27b/think_false/trace.jsonl",
   "raw/seat43/judges-nightbattery-gemma4_31b.jsonl",
   "raw/seat43/judges-nightbattery-muse-glimmer_30b.jsonl",
   "raw/seat43/judges-nightbattery-qwen3.6_27b.jsonl",
   "seat43-summary.json",
   "summary-C1-gemma4_31b.json",
   "summary-C1-muse-glimmer_30b.json",
   "summary-C1-none.json",
   "summary-C1-qwen3.6_27b.json",
   "summary-C1.json"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 2,
  "digests_written": 11,
  "digests_kept": 9,
  "chips_written": 8,
  "chips_kept": 6,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 3,
   "length": 1,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 1,
   "profanity": 0
  },
  "new_noun_tokens_checked": [
   "024-token",
   "09",
   "1",
   "11",
   "11/12",
   "12",
   "12/12",
   "13",
   "136",
   "16",
   "19",
   "23",
   "24",
   "24.0",
   "25.43",
   "27",
   "27b",
   "3090",
   "30b",
   "31b",
   "32",
   "384-token",
   "4",
   "43-case",
   "440",
   "541",
   "58Z",
   "6/9",
   "7/9",
   "768-token",
   "8",
   "9",
   "CPU",
   "GB",
   "GiB",
   "JSON",
   "MiB",
   "RANKED",
   "gemma4",
   "laguna-xs-2.1",
   "nemotron-3.5-lightning",
   "qwen3.6"
  ],
  "new_noun_tokens_withheld": 1,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "digest",
    "reason": "redaction",
    "item": null,
    "item_chars": 227,
    "withheld": "a string carrying a fenced literal",
    "detail": ""
   },
   {
    "kind": "digest",
    "reason": "figure",
    "item": "the section explains that two of five models spilled to CPU because their loaded footprints exceeded the 24 GB capacity. gemma4:31b spilled 1,541 MiB and nemotron-3.5-lightning:30b spilled 4,136 MiB.",
    "detail": "figures absent: ['30']"
   },
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "the loaded footprint for laguna-xs-2.1:latest is 19,440 MiB",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "chip",
    "reason": "figure",
    "item": "Increasing the reasoning-strength dial for muse-glimmer:30b improved its kill-recall from 11/12 to 12/12 but decreased its preservation from 7/9 to 6/9. This demonstrates a trade-off between catching ",
    "detail": "figures absent from cited spans: ['30']"
   },
   {
    "kind": "chip",
    "reason": "length",
    "item": "how does model size impact performance on a 24 GB card?",
    "detail": "q not lowercase"
   }
  ]
 },
 "generated_utc": "2026-09-09T01:26:21Z"
}