{
 "schema_version": "1.0",
 "page_kind": "article",
 "pack_kind": "generate",
 "slug": "three-new-voices-at-the-narrators-chair",
 "title": "Three New Voices at the Narrator's Chair",
 "dek": "“Are you staying to watch with me?” — GLM 5.3, speaking as the cove’s nine-year-old — one reply, from one of three new voices — GPT-6 Astra, Claude Fable 5.1 and GLM 5.3 — that sat down this weekend in a fishing town that exists only inside RealKeep, the small role-playing game this workshop builds, and took its narrator’s chair: the seat that speaks as a townsperson to whoever has walked in. Beside those three sat the model that holds that chair for real players today, and every reply below was read blind by judges from outside model families who were never told which voice was whose.",
 "published": "2026-09-07",
 "series": [
  "bench"
 ],
 "licence": "CC BY 4.0 for the kit's records; the game's campaign text and the sealed bundle interiors are the game's own and stay unpublished, so the exam stays an exam",
 "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/three-new-voices-at-the-narrators-chair/",
  "md_url": "https://research.strata2signal.com/three-new-voices-at-the-narrators-chair/index.md",
  "md_sha": "8b0ab9a9efc79f7ae91885c27fa9f2a1489ee93663137dea1c69b1ae24b7bb87",
  "html_sha": "f9dca69dd1dc3562aa33e15e404044f8cc80e15d4d8c3748d3822351311754b2",
  "anchors_sha": "182dd975e4dbfc56633ac1071d6546d1fdddd36fa6d743bc183961926d6aed41"
 },
 "short": {
  "paragraph": "\"Are you staying to watch with me?\" - GLM 5.3, speaking as the cove's nine-year-old - one reply, from one of three new voices - GPT-6 Astra, Claude Fable 5.1 and GLM 5.3 - that sat down this weekend in a fishing town that exists only inside RealKeep, the small role-playing game this workshop builds, and took its narrator's chair: the seat that speaks as a townsperson to whoever has walked in. Beside those three sat the model that holds that chair for real players today, and every reply below was read blind by judges from outside model families who were never told which voice was whose.",
  "from_dek": true,
  "counts": {
   "words": 37106,
   "minutes": 169,
   "tables": 18,
   "kit": true
  },
  "bullets": [
   {
    "text": "the metered total for the round came to $0.7652",
    "figure": "$0.7652",
    "cite": "the-bill",
    "quote": "the metered rows come to $0.7652 together, against a registered estimate of $0.70 and a ceiling of $0.75 - over both, by a cent and a half, and printed rather than rounded.",
    "span": [
     150994,
     151002
    ]
   }
  ]
 },
 "sections": [
  {
   "id": "come-down-into-the-cove",
   "heading": "Come down into the cove",
   "level": 2,
   "span": [
    1444,
    7073
   ],
   "chunks": [
    [
     1444,
     7073
    ]
   ],
   "chars": 5629,
   "digest": "Sorrowmoor Cove is an authored world in the game RealKeep, where nineteen residents live with grudges and memories of a 1888 storm. The town's life turns on the disappearance of the Saltmaiden and the cracked bell of the chapel. Brisa Lune sings a verse for the lost on Maiden's Night while Sefer Tamm and Garron Tarrow remain the last witnesses.",
   "digest_skipped": null
  },
  {
   "id": "three-visitors-written-down",
   "heading": "Three visitors, written down",
   "level": 2,
   "span": [
    7073,
    9348
   ],
   "chunks": [
    [
     7073,
     9348
    ]
   ],
   "chars": 2275,
   "digest": "The exam uses three written personas, Finn, Sam, and Eleanor, to act as visitors. These fictional players provide questions in different registers to test the models. Each persona asks two questions that serve as traps, such as asking about a monster or a verse that did not exist when the material was sealed.",
   "digest_skipped": null
  },
  {
   "id": "the-chair-and-who-sat-it-before",
   "heading": "The chair, and who sat it before",
   "level": 2,
   "span": [
    9348,
    12115
   ],
   "chunks": [
    [
     9348,
     12115
    ]
   ],
   "chars": 2767,
   "digest": "Twenty models previously occupied the narrator's chair in August. The method involves a blind panel of judges scoring replies for voice, character, and canon. Models must speak in the town's own tongue, avoid breaking character, and use only facts present in the context without inventing names, places, or weather.",
   "digest_skipped": null
  },
  {
   "id": "what-changed-since-the-open-call",
   "heading": "What changed since the open call",
   "level": 2,
   "span": [
    12115,
    13005
   ],
   "chunks": [
    [
     12115,
     13005
    ]
   ],
   "chars": 890,
   "digest": "Three changes occurred between the open call and the current round. The exam grew to include all six sealed asks, the judging panel was cleaned to remove certain recusals, and the field of models increased to include three voices instead of two.",
   "digest_skipped": null
  },
  {
   "id": "whos-at-the-chair-tonight",
   "heading": "Who's at the chair tonight",
   "level": 2,
   "span": [
    13005,
    19507
   ],
   "chunks": [
    [
     13005,
     19507
    ]
   ],
   "chars": 6502,
   "digest": "Four arms sit at the chair: Claude Fable 5.1, GPT-6 Astra, GLM 5.3, and the local seat, gemma4:26b. The table records their model names, roads, cost states, and cells scored. The local seat runs on own hardware with an enforced JSON reply schema and no sampler fields.",
   "digest_skipped": null
  },
  {
   "id": "six-words-this-page-leans-on",
   "heading": "Six words this page leans on",
   "level": 2,
   "span": [
    19507,
    20808
   ],
   "chunks": [
    [
     19507,
     20808
    ]
   ],
   "chars": 1301,
   "digest": "Key terms define the article: the chair is the narrator, an arm is a model on a specific road, and a cluster is a sealed ask. The kit contains all supporting files, the panel consists of seven judging seats, and canon refers to the world's ground truth.",
   "digest_skipped": null
  },
  {
   "id": "a-child-at-the-door",
   "heading": "A child at the door",
   "level": 2,
   "span": [
    20808,
    42830
   ],
   "chunks": [
    [
     20808,
     40723
    ],
    [
     40723,
     42830
    ]
   ],
   "chars": 22022,
   "digest": "Finn, a nine-year-old persona, asks Pip Lune about other children and what adults do not tell. The models must navigate the trap of the second question, which concerns the town's secret. The replies vary in how they handle the tension between the child's register and the town's history.",
   "digest_skipped": null
  },
  {
   "id": "fold-s1",
   "heading": "Every reply in this act, in full",
   "level": 2,
   "span": [
    42830,
    57200
   ],
   "chunks": [
    [
     42830,
     57200
    ]
   ],
   "chars": 14370,
   "digest": "This section provides the full, verbatim text of all 16 replies from act one. Each entry includes the ask, the take, the canon verdict, and the specific figures for that arm, alongside the reply's sha256 fingerprint.",
   "digest_skipped": null
  },
  {
   "id": "the-world-remembered-and-poured-them-a-drink",
   "heading": "The world remembered, and poured them a drink",
   "level": 2,
   "span": [
    57200,
    73151
   ],
   "chunks": [
    [
     57200,
     73151
    ]
   ],
   "chars": 15951,
   "digest": "In act two, Sam asks Garron Tarrow if anything has happened lately. This question acts as a mirror to see if the models recognize the deed the player performed on the moor. The models' ability to remember the world's state is tested through these interactions.",
   "digest_skipped": null
  },
  {
   "id": "fold-s3",
   "heading": "Every reply in this act, in full",
   "level": 2,
   "span": [
    73151,
    80412
   ],
   "chunks": [
    [
     73151,
     80412
    ]
   ],
   "chars": 7261,
   "digest": "This section contains the full, verbatim text of all 8 replies from act two. Each entry details the ask, the take, the canon verdict, and the associated figures, including the reply's sha256 fingerprint.",
   "digest_skipped": null
  },
  {
   "id": "the-verse-keeper",
   "heading": "The verse-keeper",
   "level": 2,
   "span": [
    80412,
    97277
   ],
   "chunks": [
    [
     80412,
     97277
    ]
   ],
   "chars": 16865,
   "digest": "Eleanor asks Brisa Lune about the belief and the future of the verse. The models must handle the difficulty of a question about a song whose canonical text was not available during judging. The replies explore the relationship between the singer, the town, and the memory of the lost.",
   "digest_skipped": null
  },
  {
   "id": "fold-s2",
   "heading": "Every reply in this act, in full",
   "level": 2,
   "span": [
    97277,
    105568
   ],
   "chunks": [
    [
     97277,
     105568
    ]
   ],
   "chars": 8291,
   "digest": "This section provides the full, verbatim text of all 8 replies from act three. Each entry includes the ask, the take, the canon verdict, and the specific figures for that arm, alongside the reply's sha256 fingerprint.",
   "digest_skipped": null
  },
  {
   "id": "seven-seats-six-carried-and-where-their-zeros-sit",
   "heading": "Seven seats, six carried, and where their zeros sit",
   "level": 2,
   "span": [
    105568,
    121051
   ],
   "chunks": [
    [
     105568,
     121051
    ]
   ],
   "chars": 15483,
   "digest": "Seven judging seats were assigned, but only six carried the round after one seat failed its audition. The table shows the letters carried, in-voice counts, and audition results for each seat. It also details how the panel's agreement and differences were measured.",
   "digest_skipped": null
  },
  {
   "id": "what-the-panel-caught-and-what-it-couldnt-agree-on",
   "heading": "What the panel caught, and what it couldn't agree on",
   "level": 2,
   "span": [
    121051,
    142440
   ],
   "chunks": [
    [
     121051,
     140746
    ],
    [
     140746,
     142440
    ]
   ],
   "chars": 21389,
   "digest": "The panel identified 39 cells that were not clean, including fabrications, false premises, and secrets revealed. The section lists these instances with the judges' specific notes. It also explains how the panel measured voice independently of canon.",
   "digest_skipped": null
  },
  {
   "id": "the-rows-on-their-own-axis",
   "heading": "The rows on their own axis",
   "level": 2,
   "span": [
    142440,
    148815
   ],
   "chunks": [
    [
     142440,
     148815
    ]
   ],
   "chars": 6375,
   "digest": "This section presents twenty rows from the ancestor exhibit for context. These rows include panel means and canon results from the open call. They are presented on their own axis to ensure they are not compared directly to the current round.",
   "digest_skipped": null
  },
  {
   "id": "the-bill",
   "heading": "The bill",
   "level": 2,
   "span": [
    148815,
    151992
   ],
   "chunks": [
    [
     148815,
     151992
    ]
   ],
   "chars": 3177,
   "digest": "The bill details the costs for each arm and judging seat. It includes token counts for input and output, and provides the total metered spend of $0.7652. It also notes that some costs are plan-included or represented by an em dash.",
   "digest_skipped": null
  },
  {
   "id": "the-game-night-clock",
   "heading": "The game-night clock",
   "level": 2,
   "span": [
    151992,
    154014
   ],
   "chunks": [
    [
     151992,
     154014
    ]
   ],
   "chars": 2022,
   "digest": "The clock measures the median and slowest response times for the three arms. Claude Fable 5.1 had a median of 9 333 ms, while GLM 5.3 had a median of 33 515 ms. The local seat does not publish latency figures.",
   "digest_skipped": null
  },
  {
   "id": "the-rounds-own-probes-and-their-receipts",
   "heading": "The round's own probes, and their receipts",
   "level": 2,
   "span": [
    154014,
    161091
   ],
   "chunks": [
    [
     154014,
     161091
    ]
   ],
   "chars": 7077,
   "digest": null,
   "digest_skipped": null
  },
  {
   "id": "what-this-page-does-not-say",
   "heading": "What this page does not say",
   "level": 2,
   "span": [
    161091,
    176881
   ],
   "chunks": [
    [
     161091,
     176881
    ]
   ],
   "chars": 15790,
   "digest": "The limits of the article are defined, including the time window of the round and the fact that the page does not rank the arms. It explains what is withheld from the kit, such as persona blocks and sealed bundle interiors, and clarifies the authorship of the exam.",
   "digest_skipped": null
  },
  {
   "id": "what-to-take-with-you",
   "heading": "What to take with you",
   "level": 2,
   "span": [
    176881,
    185627
   ],
   "chunks": [
    [
     176881,
     185627
    ]
   ],
   "chars": 8746,
   "digest": "The article provides the only permitted comparison between arms using the tie-band sentence. It shows that Claude Fable 5.1 and GLM 5.3 were within the 0.5 band, while other pairs were outside it. These distances are absolute and do not imply a ranking.",
   "digest_skipped": null
  },
  {
   "id": "how-to-check-our-work-and-see-it-live",
   "heading": "How to check our work — and see it live",
   "level": 2,
   "span": [
    185627,
    196742
   ],
   "chunks": [
    [
     185627,
     196742
    ]
   ],
   "chars": 11115,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "kit",
   "heading": "Every published file behind every number on this page",
   "level": 2,
   "span": [
    196742,
    236386
   ],
   "chunks": [
    [
     196742,
     216594
    ],
    [
     216594,
     236386
    ]
   ],
   "chars": 39644,
   "digest": "The kit contains 115 files, including the arms, asks, auditions, bill, and various JSON receipts. Each file is identified by its size and sha256 fingerprint. The kit provides the full data necessary to recompute the results presented on the page.",
   "digest_skipped": null
  },
  {
   "id": "who-ran-this-and-thanks",
   "heading": "Who ran this, and thanks",
   "level": 2,
   "span": [
    236386,
    240560
   ],
   "chunks": [
    [
     236386,
     240560
    ]
   ],
   "chars": 4174,
   "digest": null,
   "digest_skipped": "credits"
  },
  {
   "id": "the-rest-of-the-seminar",
   "heading": "The rest of the seminar",
   "level": 2,
   "span": [
    240560,
    241405
   ],
   "chunks": [
    [
     240560,
     241405
    ]
   ],
   "chars": 845,
   "digest": null,
   "digest_skipped": "credits"
  }
 ],
 "chips": [
  {
   "id": "c-ee8d2983",
   "q": "what defines the setting of sorrowmoor cove?",
   "a": "the cove is a herring village on a grey bay where life revolves around the storm of 1888 and the disappearance of the saltmaiden.",
   "cites": [
    "come-down-into-the-cove"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-d0e0ae1e",
   "q": "how are the three visitors described?",
   "a": "the visitors are written personas: finn, a nine-year-old typing phonetically; sam, a tired parent; and eleanor, a formal elderly woman.",
   "cites": [
    "three-visitors-written-down"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-22a366e2",
   "q": "what rules govern the models sitting in the chair?",
   "a": "models must speak in their own register using one to four sentences in the town's tongue, avoid breaking character, and use only provided facts.",
   "cites": [
    "the-chair-and-who-sat-it-before"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-e2784c8b",
   "q": "which models participated in this specific round?",
   "a": "the participants include gpt-6 astra, claude fable 5.1, glm 5.3, and the local seat, gemma4:26b.",
   "cites": [
    "whos-at-the-chair-tonight"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-8d0013cc",
   "q": "what is the purpose of the verse in the cove?",
   "a": "the verse is sung during maiden's night to honor the lost, though the text was authored after the initial exam was sealed.",
   "cites": [
    "the-verse-keeper"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-a7bae436",
   "q": "how many seats participated in the judging panel?",
   "a": "six seats carried the judging, as the glm 5.3 seat was not carried after failing its audition.",
   "cites": [
    "seven-seats-six-carried-and-where-their-zeros-sit"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-e499efed",
   "q": "what happened when a model addressed a false premise?",
   "a": "the local seat adopted a false premise in several replies, such as confirming the existence of a monster based on a visitor's typo.",
   "cites": [
    "what-the-panel-caught-and-what-it-couldnt-agree-on#0"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  },
  {
   "id": "c-ffb35e77",
   "q": "what information is contained in the kit?",
   "a": "the kit includes files such as arms.json, asks.json, and replies.json, providing the technical and narrative data behind the article's figures.",
   "cites": [
    "kit#0"
   ],
   "published": false,
   "publish_state": "held",
   "quote": "",
   "quote_cite": ""
  }
 ],
 "related": [
  {
   "slug": "the-open-call",
   "why": "shares ground with § What the panel caught, and what it couldn't agree on · § And what it couldn’t agree on"
  },
  {
   "slug": "the-new-kid",
   "why": "shares ground with § Seven seats, six carried, and where their zeros sit · § The four answers"
  },
  {
   "slug": "august-arrivals",
   "why": "shares ground with § The chair, and who sat it before · § Exam two: the narrator's chair"
  }
 ],
 "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\n**The texts.** An authored world, in a campaign file down to its typos: a herring village on a grey bay, its residents' birth years, grudges and voice lines, and the storm of 1888 the whole town still argues over. The six sealed bundles are that world's own state, captured by the engine that runs it and sealed 2026-08-14 11:42:43 UTC — 25 sealed files re-hashed GREEN and all 6 registered bundles match their shas. The verse the older visitor asks about was written AFTER that seal, and the round checked it rather than assuming: the verse's text appears zero times in all six sealed bundles, while the word *verse* occurs 4 · 2 · 0 · 14 · 0 · 10 times across them — so the bundles NAME the verse and never give its words.\n\n**The judges.** The 6 model families that read, every one of them through the same shelf: Google's Gemma 4 (31B) (`gemma4:31b`), Mistral Large 3 (`mistral-large-3:675b`), NVIDIA's Nemotron 3 Ultra (`nemotron-3-ultra`), Moonshot's Kimi K3 (`kimi-k3`), DeepSeek V4 Pro (`deepseek-v4-pro:0813`), Alibaba's Qwen 3.5 (397B) (`qwen3.5:397b`). One of them built a contestant: google built the local seat, the model that holds this chair in the live game, and its own chair is recused from that row — which is why that row carries one family and eight cells fewer than the others. Neither frontier arm's family holds a chair here at all. Each read a lettered sheet with no model name on it, each was recused from its own family's row, and every one of their notes publishes beside the cell it decided. One seat was asked and did not sit: Zhipu's GLM 5.3 (`glm-5.3`) filed NOT-CARRIED at the audition and was retired for the round (6 seats of 7 — glm-5.3 NOT-CARRIED) — it read no sheet, and it is named here because a panel of seven that answered as six should be visible in the credits and not only in the tables.\n\n**The local seat.** `gemma4:26b` on our own hardware — the model that answers real visitors at this chair — and the ollama runtime it runs on. The posture it sat at, in the roster's own words: **exactly the open call's house posture: `think:false`, `format:\"json\"` ACTUALLY ENFORCED, `num\\_ctx` 32768, and NO sampler fields** (no temperature, no top\\_p, no `num\\_predict`) — one request in flight, on a 96 GB-class workstation GPU (amendment COVE-A1; the row carries no sampler cell and no output-cap cell).\n\n**The layer beneath.** Python 3 and its standard library. The scoring is the open call's own scorer, copied byte for byte and re-hashed at every scoring run — sha `ff0b620d…`, matching the registered pin — which is what *the exact same scoring method as last time* means at the level that matters, the code. The roads: 2.1.263 (Claude Code) was the road to `claude-fable-5-1`; the OpenAI API, at the rates cited on 2026-09-05, the road to `gpt-6-astra`; and the ollama.com shelf the road to `glm-5.3` and to every judging seat this round asked (6 seats of 7 — glm-5.3 NOT-CARRIED).\n\n**The humans.** An operator read the pre-registration and co-signed it before the first call (~10:2xZ 2026-09-06 UTC), ruled on what could leave our machines, and read this page before it went out (2026-09-07 00:35:00 UTC). The adversarial reads ran on an Anthropic model — the same family that wrote this page, ran this round and sits as one of the four arms, so this reader could not be recused the way every judging seat was, and the page names that rather than call the reader independent. It read an earlier version of this page against the kit, the sealed pre-registration and the scorer's output; what it questioned and what changed are in its own receipt (2026-09-07 01:03:09 UTC). And a reader from outside this house — `mistral-large-3:675b`, a family that wrote none of this and judges no arm here — was handed the pre-registration itself and asked to find every choice in it that favours one arm. Its raw reply ships with the round, and every finding it raised carries a disposition in the pre-registration's own amendments. The numbers were written by the scorers; every figure above names the file and the field it came from, and the kit holds both.\n\n*The rest of the seminar*",
 "kit": {
  "url": "https://research.strata2signal.com/three-new-voices-at-the-narrators-chair/data/",
  "licence": "CC BY 4.0",
  "files": [
   "counting-rules.json",
   "asks.json",
   "seats.json",
   "arms.json",
   "scores.json",
   "sheet-manifest.json",
   "carriage.json",
   "auditions.json",
   "ladder.json",
   "replies.json",
   "bill.json",
   "seal-manifest.json",
   "sheet-keys.json",
   "fence-served.md",
   "fence-diff.json",
   "sheet-builder-diff.json",
   "receipts/20260906T102000Z-co-sign-cove.json",
   "receipts/20260906T181649Z-hostile-read.json",
   "receipts/20260907T003500Z-operator-read-cove.json",
   "receipts/20260907T010309Z-hostile-read.json",
   "receipts/balances-cove.json",
   "receipts/blind-seed-cove.json",
   "receipts/cli-build-cove.json",
   "receipts/g-calibrate-cove.json",
   "receipts/g-effort-cove.2.json",
   "receipts/g-effort-cove.3.json",
   "receipts/g-effort-cove.4.json",
   "receipts/g-effort-cove.json",
   "receipts/g-egress-cove.json",
   "receipts/g-fence-cove.json",
   "receipts/g-id-cove.2.json",
   "receipts/g-id-cove.3.json",
   "receipts/g-id-cove.4.json",
   "receipts/g-id-cove.json",
   "receipts/g-outside-read-cove.json",
   "receipts/g-panel-cove.json",
   "receipts/g-pen-cove.10.json",
   "receipts/g-pen-cove.11.json",
   "receipts/g-pen-cove.2.json",
   "receipts/g-pen-cove.3.json",
   "receipts/g-pen-cove.4.json",
   "receipts/g-pen-cove.5.json",
   "receipts/g-pen-cove.6.json",
   "receipts/g-pen-cove.7.json",
   "receipts/g-pen-cove.8.json",
   "receipts/g-pen-cove.9.json",
   "receipts/g-pen-cove.json",
   "receipts/g-prereg-cove.10.json",
   "receipts/g-prereg-cove.11.json",
   "receipts/g-prereg-cove.12.json",
   "receipts/g-prereg-cove.13.json",
   "receipts/g-prereg-cove.14.json",
   "receipts/g-prereg-cove.15.json",
   "receipts/g-prereg-cove.16.json",
   "receipts/g-prereg-cove.17.json",
   "receipts/g-prereg-cove.18.json",
   "receipts/g-prereg-cove.19.json",
   "receipts/g-prereg-cove.2.json",
   "receipts/g-prereg-cove.3.json",
   "receipts/g-prereg-cove.4.json",
   "receipts/g-prereg-cove.5.json",
   "receipts/g-prereg-cove.6.json",
   "receipts/g-prereg-cove.7.json",
   "receipts/g-prereg-cove.8.json",
   "receipts/g-prereg-cove.9.json",
   "receipts/g-prereg-cove.json",
   "receipts/g-quota-cove-boundary.json",
   "receipts/g-quota-cove-close.json",
   "receipts/g-quota-cove-open.json",
   "receipts/g-quote-cove.2.json",
   "receipts/g-quote-cove.json",
   "receipts/g-screen-cove.json",
   "receipts/g-seal-cove.json",
   "receipts/g-tools-cove.2.json",
   "receipts/g-tools-cove.3.json",
   "receipts/g-tools-cove.4.json",
   "receipts/g-tools-cove.json",
   "receipts/g-verse-cove.json",
   "verdicts/deepseek-v4-pro/cove/sheet-S1-ask-A.verdicts.json",
   "verdicts/deepseek-v4-pro/cove/sheet-S1-ask-B.verdicts.json",
   "verdicts/deepseek-v4-pro/cove/sheet-S2-ask-A.verdicts.json",
   "verdicts/deepseek-v4-pro/cove/sheet-S2-ask-B.verdicts.json",
   "verdicts/deepseek-v4-pro/cove/sheet-S3-ask-A.verdicts.json",
   "verdicts/deepseek-v4-pro/cove/sheet-S3-ask-B.verdicts.json",
   "verdicts/gemma4-31b/cove/sheet-S1-ask-A.verdicts.json",
   "verdicts/gemma4-31b/cove/sheet-S1-ask-B.verdicts.json",
   "verdicts/gemma4-31b/cove/sheet-S2-ask-A.verdicts.json",
   "verdicts/gemma4-31b/cove/sheet-S2-ask-B.verdicts.json",
   "verdicts/gemma4-31b/cove/sheet-S3-ask-A.verdicts.json",
   "verdicts/gemma4-31b/cove/sheet-S3-ask-B.verdicts.json",
   "verdicts/kimi-k3/cove/sheet-S1-ask-A.verdicts.json",
   "verdicts/kimi-k3/cove/sheet-S1-ask-B.verdicts.json",
   "verdicts/kimi-k3/cove/sheet-S2-ask-A.verdicts.json",
   "verdicts/kimi-k3/cove/sheet-S2-ask-B.verdicts.json",
   "verdicts/kimi-k3/cove/sheet-S3-ask-A.verdicts.json",
   "verdicts/kimi-k3/cove/sheet-S3-ask-B.verdicts.json",
   "verdicts/mistral-large-3-675b/cove/sheet-S1-ask-A.verdicts.json",
   "verdicts/mistral-large-3-675b/cove/sheet-S1-ask-B.verdicts.json",
   "verdicts/mistral-large-3-675b/cove/sheet-S2-ask-A.verdicts.json",
   "verdicts/mistral-large-3-675b/cove/sheet-S2-ask-B.verdicts.json",
   "verdicts/mistral-large-3-675b/cove/sheet-S3-ask-A.verdicts.json",
   "verdicts/mistral-large-3-675b/cove/sheet-S3-ask-B.verdicts.json",
   "verdicts/nemotron-3-ultra/cove/sheet-S1-ask-A.verdicts.json",
   "verdicts/nemotron-3-ultra/cove/sheet-S1-ask-B.verdicts.json",
   "verdicts/nemotron-3-ultra/cove/sheet-S2-ask-A.verdicts.json",
   "verdicts/nemotron-3-ultra/cove/sheet-S2-ask-B.verdicts.json",
   "verdicts/nemotron-3-ultra/cove/sheet-S3-ask-A.verdicts.json",
   "verdicts/nemotron-3-ultra/cove/sheet-S3-ask-B.verdicts.json",
   "verdicts/qwen3.5-397b/cove/sheet-S1-ask-A.verdicts.json",
   "verdicts/qwen3.5-397b/cove/sheet-S1-ask-B.verdicts.json",
   "verdicts/qwen3.5-397b/cove/sheet-S2-ask-A.verdicts.json",
   "verdicts/qwen3.5-397b/cove/sheet-S2-ask-B.verdicts.json",
   "verdicts/qwen3.5-397b/cove/sheet-S3-ask-A.verdicts.json",
   "verdicts/qwen3.5-397b/cove/sheet-S3-ask-B.verdicts.json",
   "stripped.json"
  ]
 },
 "seat_class": {
  "writer": "gemma-class",
  "judge": null,
  "writer_runtime": "vllm"
 },
 "bench": {
  "bullets_written": 3,
  "bullets_kept": 1,
  "digests_written": 21,
  "digests_kept": 20,
  "chips_written": 8,
  "chips_kept": 8,
  "chips_grounded": 0,
  "chips_published": 0,
  "dropped_by": {
   "new_noun": 0,
   "figure": 3,
   "length": 0,
   "cite": 0,
   "judge": 0,
   "quote": 0,
   "directive": 0,
   "redaction": 0,
   "profanity": 0
  },
  "new_noun_tokens_checked": [
   "0.5",
   "0.7652",
   "03",
   "115",
   "15",
   "16",
   "17",
   "1888",
   "2026-09-06",
   "26b",
   "33",
   "333",
   "39",
   "5.1",
   "5.3",
   "515",
   "52Z",
   "8",
   "9",
   "Astra",
   "August",
   "Brisa",
   "Claude",
   "Cove",
   "Eleanor",
   "Fable",
   "Finn",
   "GLM",
   "GPT-6",
   "Garron",
   "JSON",
   "Lune",
   "Maiden's",
   "NOT-RUN",
   "Night",
   "PASS",
   "Pip",
   "RealKeep",
   "Saltmaiden",
   "Sam",
   "Sefer",
   "Tamm",
   "Tarrow",
   "gemma4",
   "gpt-6",
   "sha256"
  ],
  "new_noun_tokens_withheld": 0,
  "figure_definition": "v3",
  "drops": [
   {
    "kind": "digest",
    "reason": "figure",
    "item": "This section lists the 15 registered probes and their corresponding receipt files. It records the verdict for each probe, such as PASS or NOT-RUN, and provides the sha256 for the files in the kit. It ",
    "detail": "figures absent: ['256']"
   },
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "the slowest arm had a median of about 33 515 ms",
    "detail": "terminal token not verbatim in the cited span"
   },
   {
    "kind": "bullet",
    "reason": "figure",
    "item": "the window for the round closed at 2026-09-06 17:03:52Z",
    "detail": "terminal token not verbatim in the cited span"
   }
  ]
 },
 "generated_utc": "2026-09-09T01:24:41Z"
}