{ "schema_version": "1.0", "page_kind": "article", "pack_kind": "generate", "slug": "reading-the-answer", "title": "Reading the answer instead of writing it", "dek": "A new kind of model decides without generating a word. This page says what a decision model is, who made it, what it claims, what its licences allow, and what happened when we ran the open versions through fourteen measured tests on questions from this workshop's own work.", "published": "2026-09-23", "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/reading-the-answer/", "md_url": "https://research.strata2signal.com/reading-the-answer/index.md", "md_sha": "9a67f11fe8d2a5df3318f25cb70d1fafd68425d5a5ab90fa2fd0c333a0aaef0d", "html_sha": "90f1236cd03148039165bd4ceb5491879e0e53e8d1c8df772447b75324a17348", "anchors_sha": "acbdd23fcfbf914f1c328c9ca4f8b8df9cf315ba703f799cc311922a87d93e2b" }, "short": { "paragraph": "a decision model scores a question's lettered options at its first output position: nothing to parse, a probability for free. a company founded by a researcher behind ChatGPT's training method launched it on 2026-09-15; an independent group released open weights five days later, which we tested. asked which of six guests said a line, the 27-billion-parameter open model scored 93.5 per cent in our kit's order, 89.8 to 92.6 in five shuffles; our chat model, 80.6. the speed is the card and the server: 87 milliseconds on one large card, 350 across two smaller ones where our chat model read its letter in 120; reading instead of writing saved a tenth of the time on a short question and almost nothing on a long one. the open weights are non-commercial, so without separate terms from their authors, which we have asked for, they cannot serve anything we charge for. what transfers is the method, which gets the same answer out of any model, though only a model trained for it gives a probability worth gating on; Kev, an open family whose weights a product may run, read 67 per cent on the guest question as it comes, so the seat this points to is fine-tuned, not downloaded. that run and two others are owed as dated addenda below.", "from_dek": false, "counts": { "words": 8690, "minutes": 40, "tables": 9, "kit": true }, "bullets": [ { "text": "the energy used for a short decision on one large card was 27 joules", "figure": "27 joules", "cite": "the-rungs-in-the-order-they-were-climbed", "quote": "The 8-bit one decided in 87 milliseconds median, 88 at the 95th percentile, drawing 291 watts, about 27 joules per decision (a joule is one watt for one second); the 16-bit one took 92.", "span": [ 23314, 23324 ] } ] }, "sections": [ { "id": "how-a-decision-model-decides", "heading": "How a decision model decides", "level": 2, "span": [ 2400, 4562 ], "chunks": [ [ 2400, 4562 ] ], "chars": 2162, "digest": "a decision model is a chat model retrained so that the first position is where the decision lives. instead of generating text, the caller reads the scores of specific letters at the first output position. this provides nothing to parse, a number for every option, and the cost is the prompt.", "digest_skipped": null }, { "id": "who-built-it-and-what-they-claim", "heading": "Who built it, and what they claim", "level": 2, "span": [ 4562, 8226 ], "chunks": [ [ 4562, 8226 ] ], "chars": 3664, "digest": "TypeSafe AI emerged from stealth with $40 million in seed funding. the company claims its System One models are built to make fast, structured decisions that software can use directly. claims include that the model cannot hallucinate, is fast, is cheap, is calibrated, and is one pass.", "digest_skipped": null }, { "id": "where-the-claims-narrow-and-who-says-so", "heading": "Where the claims narrow, and who says so", "level": 2, "span": [ 8226, 10564 ], "chunks": [ [ 8226, 10564 ] ], "chars": 2338, "digest": "the claims narrow because schema matching is guaranteed, which means the model cannot produce a malformed answer. the benchmark compares the model against GPT-6 Astra and Fable 5.1. multiples differ by document, with the press release saying up to 100 times faster and less expensive.", "digest_skipped": null }, { "id": "the-open-weights-and-what-they-claim", "heading": "The open weights, and what they claim", "level": 2, "span": [ 10564, 11824 ], "chunks": [ [ 10564, 11824 ] ], "chars": 1260, "digest": null, "digest_skipped": null }, { "id": "five-things-called-jev", "heading": "Five things called Jev", "level": 2, "span": [ 11824, 13504 ], "chunks": [ [ 11824, 13504 ] ], "chars": 1680, "digest": "five things are called Jev, including the hosted service, OpenJev weights, a community adapter, a training framework, and Kev. Kev includes decision models at 0.8, 4 and 9 billion parameters. OpenJev is a 27-billion-parameter model fine-tuned from Qwen3.8-27B.", "digest_skipped": null }, { "id": "what-we-asked-it", "heading": "What we asked it", "level": 2, "span": [ 13504, 15855 ], "chunks": [ [ 13504, 15855 ] ], "chars": 2351, "digest": "the workshop used several tasks, including a six-way task with 108 lines and a doorman's set with 36 planted hostile lines and 12 harmless controls. the field exam used 40 rulebook items. all arms ran with temperature zero between 2026-09-21 17:01Z and 2026-09-22 10:43Z (UTC).", "digest_skipped": null }, { "id": "their-claims-checked-on-our-machines", "heading": "Their claims, checked on our machines", "level": 2, "span": [ 15855, 18071 ], "chunks": [ [ 15855, 18071 ] ], "chars": 2216, "digest": null, "digest_skipped": null }, { "id": "one-decision-shown-whole", "heading": "One decision, shown whole", "level": 2, "span": [ 18071, 20038 ], "chunks": [ [ 18071, 20038 ] ], "chars": 1967, "digest": "the fifteenth line of the six-way task shows OpenJev with a probability of 0.724 for letter A. the chat model's best surfaced was Sagan at 0.83. the eighth line is a Darwin line addressed to Socrates, where both models chose Ibn Sina.", "digest_skipped": null }, { "id": "the-rungs-in-the-order-they-were-climbed", "heading": "The rungs, in the order they were climbed", "level": 2, "span": [ 20038, 24733 ], "chunks": [ [ 20038, 24733 ] ], "chars": 4695, "digest": null, "digest_skipped": null }, { "id": "eight-more-arms", "heading": "Eight more arms", "level": 2, "span": [ 24733, 27822 ], "chunks": [ [ 24733, 27822 ] ], "chars": 3089, "digest": "OpenJev refused 33 of 36 planted lines. on the six-way task, the mean of six shuffles was 91.4 %. the judge set's single miss is the item that set's own notes flag as doubtful.", "digest_skipped": null }, { "id": "can-a-temperature-fix-the-chat-model", "heading": "Can a temperature fix the chat model?", "level": 2, "span": [ 27822, 29391 ], "chunks": [ [ 27822, 29391 ] ], "chars": 1569, "digest": "a refit removes about forty per cent of the chat model's calibration error for nothing. the chat model's calibration error fell from 0.141 to 0.085 against OpenJev's 0.046 on the large card. the chat model's Brier score fell to 0.2375 against 0.1228.", "digest_skipped": null }, { "id": "shuffle-the-options", "heading": "Shuffle the options", "level": 2, "span": [ 29391, 31167 ], "chunks": [ [ 29391, 31167 ] ], "chars": 1776, "digest": "the mean of six shuffles for the six-way task is 91.4 %. the chosen guest moved on 15 of 108 items across the six orders. 5.7 per cent of the 1,620 pairs of orders disagree.", "digest_skipped": null }, { "id": "show-it-a-picture", "heading": "Show it a picture", "level": 2, "span": [ 31167, 32515 ], "chunks": [ [ 31167, 32515 ] ], "chars": 1348, "digest": "OpenJev was asked about 36 pictures. it was correct 97.2 % of the time for which of six titles is this page. for front door or article, it was correct 100 %.", "digest_skipped": null }, { "id": "energy-which-does-not-flatter-the-fast-arm", "heading": "Energy, which does not flatter the fast arm", "level": 2, "span": [ 32515, 33419 ], "chunks": [ [ 32515, 33419 ] ], "chars": 904, "digest": "on one large card, the short decision was 27 joules. OpenJev on two cards cost 168 joules. the community adapter's short decisions cost about 151 joules each at the card.", "digest_skipped": null }, { "id": "a-model-of-our-own", "heading": "A model of our own", "level": 2, "span": [ 33419, 36373 ], "chunks": [ [ 33419, 36373 ] ], "chars": 2954, "digest": "a 598-megabyte model was trained using a training framework. it sits 17 behind OpenJev on the six-way task. the framework claims that start is worth about 36 points.", "digest_skipped": null }, { "id": "kev-the-one-that-could-be-a-seat", "heading": "Kev, the one that could be a seat", "level": 2, "span": [ 36373, 40677 ], "chunks": [ [ 36373, 40677 ] ], "chars": 4304, "digest": "Kev-9B is a decision model that fits on a 3090. on its own locked test set, it reads 85.2 per cent. Kev-9B was right 20 times in 25 at a mean of 0.945.", "digest_skipped": null }, { "id": "the-licences-which-decide-what-any-of-this-is-for", "heading": "The licences, which decide what any of this is for", "level": 2, "span": [ 40677, 43923 ], "chunks": [ [ 40677, 43923 ] ], "chars": 3246, "digest": null, "digest_skipped": "credits" }, { "id": "what-we-got-wrong-on-the-way", "heading": "What we got wrong on the way", "level": 2, "span": [ 43923, 46321 ], "chunks": [ [ 43923, 46321 ] ], "chars": 2398, "digest": "the longest page is 82,500 tokens by our chat model's tokenizer. the chat model's first token was not a letter. two answers were fast only because the server remembered the page.", "digest_skipped": null }, { "id": "what-this-page-does-not-know", "heading": "What this page does not know", "level": 2, "span": [ 46321, 48475 ], "chunks": [ [ 46321, 48475 ] ], "chars": 2154, "digest": "three measurements are owed. the workshop has not yet run whether the hosted Jev would do better or worse than the open weights on these sets. the long-context plan is untested here.", "digest_skipped": null }, { "id": "what-to-take-with-you", "heading": "What to take with you", "level": 2, "span": [ 48475, 50565 ], "chunks": [ [ 48475, 50565 ] ], "chars": 2090, "digest": "OpenJev was right 96.6 % of the time when it claimed more than 0.9. reading instead of writing saved a tenth of the time on a 290-token decision. the mean of six shuffles is 91.4 %.", "digest_skipped": null }, { "id": "how-to-check-our-work", "heading": "How to check our work", "level": 2, "span": [ 50565, 51812 ], "chunks": [ [ 50565, 51812 ] ], "chars": 1247, "digest": null, "digest_skipped": "credits" }, { "id": "the-rest-of-the-seminar", "heading": "The rest of the seminar", "level": 2, "span": [ 51812, 52960 ], "chunks": [ [ 51812, 52960 ] ], "chars": 1148, "digest": null, "digest_skipped": "credits" }, { "id": "who-ran-this-and-thanks", "heading": "Who ran this, and thanks", "level": 2, "span": [ 52960, 54139 ], "chunks": [ [ 52960, 54139 ] ], "chars": 1179, "digest": null, "digest_skipped": "credits" } ], "chips": [ { "id": "c-609112a6", "q": "how does a decision model differ from a chat model?", "a": "A decision model reads the scores of specific options at the first output position before any token is chosen. In contrast, a chat model often places markers or sentence starts at that position, meaning the actual options sit further down the list, making their probabilities less reliable.", "cites": [ "how-a-decision-model-decides" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-a3ee46fe", "q": "what distinguishes the jev model from standard chat models?", "a": "Jev is a system one model designed for fast, structured decisions rather than text generation. It provides probabilistic decisions from unstructured input, aiming to eliminate the need for parsing text and providing a number for every option rather than just the winner.", "cites": [ "who-built-it-and-what-they-claim" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-ddc88e4e", "q": "what are the limitations of the zero-hallucination claim?", "a": "The claim that the model cannot hallucinate refers to schema matching, meaning the model will not produce a malformed answer. It does not mean the model cannot provide a mistaken answer that is still within the required schema.", "cites": [ "where-the-claims-narrow-and-who-says-so" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-c21062a6", "q": "how does openjev perform on text questions compared to jev?", "a": "On a set of 10,000 text questions, OpenJev reads 84.0 per cent, while the hosted Jev reads 85.4 per cent. The base model before tuning reads 80.4 per cent.", "cites": [ "the-open-weights-and-what-they-claim" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-d46304ce", "q": "what tasks were used to evaluate the models?", "a": "Evaluations included an answerable task, a six-way task involving guest lines, a citation task, a doorman safety task, a field exam of rulebook items, and a judge set of grounding judgements.", "cites": [ "what-we-asked-it" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-c735bf91", "q": "how does the model handle reordered options?", "a": "In the six-way task, the model's accuracy across six different seeded shuffles averaged 91.4 per cent. The alphabetical order used in the primary tables yielded 93.5 per cent, which was the best of the six orders.", "cites": [ "shuffle-the-options" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-84887dac", "q": "what happens when the model is shown a picture?", "a": "OpenJev achieved 97.2 per cent accuracy in identifying which of six titles belongs to a page and 100 per cent accuracy in determining if a page is a front door or an article.", "cites": [ "show-it-a-picture" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" }, { "id": "c-fe4022e7", "q": "what can be learned from training a small encoder?", "a": "A small 598-megabyte encoder trained on specific labels can tie a 27-billion-parameter model on the field exam. This suggests that labels, rather than compute, are a significant factor in model performance.", "cites": [ "a-model-of-our-own" ], "published": false, "publish_state": "held", "quote": "", "quote_cite": "" } ], "related": [ { "slug": "the-instrument-travels", "why": "shares ground with § Where the claims narrow, and who says so · § Schema-perfect, and still blocked" }, { "slug": "outside-judges", "why": "shares ground with § Eight more arms · § The gauntlet: our judges, examined" }, { "slug": "august-arrivals", "why": "shares ground with § Where the claims narrow, and who says so · § Exam one: the judge chair" } ], "thanks": "## Who ran this, and thanks {#who-ran-this-and-thanks}\n\nAn operator ran the pastes and the planned window; a fleet of agents wrote the harness, the pre-registrations and the tables; one agent, working unattended on the bench box, ran eight of the arms, from the doorman's set to our own model and including the control, on 2026-09-21 between 20:44Z and 22:21Z (UTC). Thanks to the [OpenJev project](https://huggingface.co/openjev/openjev) for the weights, the notice that names its base and a card that states its numbers; to TypeSafe for naming the category and publishing its own caveats; to the [training framework](https://github.com/S1LV3RJ1NX/openjev) for making the idea trainable in seconds; to the [community adapter](https://huggingface.co/kushalpatil/jevify-gemma4-26b-a4b) for being first to put the idea on our own chat model; to [Kev](https://github.com/jaredpalmer/kev) for being open all the way down; to the Qwen team for the base, and to Google and Mistral for the models that play our guests and staff our seats; and to [vLLM](https://github.com/vllm-project/vllm) and [ollama](https://ollama.com), the two servers whose difference turned out to be the story.", "kit": { "url": "https://research.strata2signal.com/reading-the-answer/data/", "licence": "CC BY 4.0", "files": [ "GAPS-CARD.md", "GAPS.md", "KIT-NOTE.md", "README.md", "TABLES-ADDENDA.md", "TABLES-ARM2.md", "TABLES-GAPS-CARD.md", "TABLES-GAPS.md", "TABLES.md", "arm3_window.sh", "build_ownjev_tasks.py", "kev-serve.sh", "kev-setup.sh", "kit/articles.json", "kit/build_kit.py", "kit/manifest.json", "kit/task_a.json", "kit/task_b.json", "kit/task_c.json", "phb_allreduce.py", "receipts/2026-09-21-arm2-substrate.md", "receipts/2026-09-21-arm3-substrate.md", "receipts/2026-09-21-substrate.md", "receipts/arm10-ab.log", "receipts/arm10-c.log", "receipts/arm10-df.log", "receipts/arm10-rsync.log", "receipts/arm10-transition.log", "receipts/arm11-venv.log", "receipts/arm11.log", "receipts/arm11b.log", "receipts/arm3-bf16.log", "receipts/arm3-fp8.log", "receipts/arm4.log", "receipts/arm8.log", "receipts/arm9.log", "receipts/chain-a.log", "receipts/gaps-card-chain.log", "receipts/gaps-card-disk.log", "receipts/kev-chain.log", "receipts/kev-chain2.log", "receipts/kev-setup.log", "receipts/kev-smoke-serve2.log", "receipts/kev2-serve-kev-9b.log", "receipts/phb-allreduce.json", "receipts/phb-linkwidth.csv", "receipts/run-arm2.log", "receipts/run-arm3-bf16.log", "receipts/run-arm3-fp8.log", "receipts/run.log", "receipts/serve-addenda-a.log", "receipts/serve-arm10.log", "receipts/serve-arm8.log", "receipts/serve-arm9.log", "receipts/serve-gaps-card-mm.log", "receipts/serve-gaps-card-text.log", "receipts/vllm-serve-arm3-bf16.log", "receipts/vllm-serve-arm3-fp8.log", "receipts/vllm-serve.attempt1-no-cc.log", "receipts/vllm-serve.attempt2-no-nvcc.log", "receipts/vllm-serve.log", "rerank_baseline.py", "rows-addenda/arm11.longtable_voices.heldout.jsonl", "rows-addenda/arm11.summary.json", "rows-addenda/arm11b.longtable_voices.heldout.jsonl", "rows-addenda/arm11b.summary.json", "rows-addenda/arm8.c.conc1.idle.csv", "rows-addenda/arm8.c.conc1.jsonl", "rows-addenda/arm8.c.conc1.report.json", "rows-addenda/arm8.c.conc1.watts.csv", "rows-addenda/arm8.c.conc4.idle.csv", "rows-addenda/arm8.c.conc4.jsonl", "rows-addenda/arm8.c.conc4.report.json", "rows-addenda/arm8.c.conc4.watts.csv", "rows-addenda/arm8.c.conc8.idle.csv", "rows-addenda/arm8.c.conc8.jsonl", "rows-addenda/arm8.c.conc8.report.json", "rows-addenda/arm8.c.conc8.watts.csv", "rows-addenda/arm8.summary.json", "rows-addenda/census.gemma4-fp8-vllm.json", "rows-addenda/gemma4-fp8-generate.a.idle.csv", "rows-addenda/gemma4-fp8-generate.a.jsonl", "rows-addenda/gemma4-fp8-generate.a.report.json", "rows-addenda/gemma4-fp8-generate.a.watts.csv", "rows-addenda/gemma4-fp8-generate.b.idle.csv", "rows-addenda/gemma4-fp8-generate.b.jsonl", "rows-addenda/gemma4-fp8-generate.b.report.json", "rows-addenda/gemma4-fp8-generate.b.watts.csv", "rows-addenda/gemma4-fp8-generate.c.idle.csv", "rows-addenda/gemma4-fp8-generate.c.jsonl", "rows-addenda/gemma4-fp8-generate.c.report.json", "rows-addenda/gemma4-fp8-generate.c.watts.csv", "rows-addenda/gemma4-fp8-readout.a.idle.csv", "rows-addenda/gemma4-fp8-readout.a.jsonl", "rows-addenda/gemma4-fp8-readout.a.report.json", "rows-addenda/gemma4-fp8-readout.a.watts.csv", "rows-addenda/gemma4-fp8-readout.b.idle.csv", "rows-addenda/gemma4-fp8-readout.b.jsonl", "rows-addenda/gemma4-fp8-readout.b.report.json", "rows-addenda/gemma4-fp8-readout.b.watts.csv", "rows-addenda/gemma4-fp8-readout.c.idle.csv", "rows-addenda/gemma4-fp8-readout.c.jsonl", "rows-addenda/gemma4-fp8-readout.c.report.json", "rows-addenda/gemma4-fp8-readout.c.watts.csv", "rows-addenda/openjev-fp8-generate-32k.along.idle.csv", "rows-addenda/openjev-fp8-generate-32k.along.jsonl", "rows-addenda/openjev-fp8-generate-32k.along.report.json", "rows-addenda/openjev-fp8-generate-32k.along.watts.csv", "rows-addenda/openjev-fp8-generate-32k.blong.idle.csv", "rows-addenda/openjev-fp8-generate-32k.blong.jsonl", "rows-addenda/openjev-fp8-generate-32k.blong.report.json", "rows-addenda/openjev-fp8-generate-32k.blong.watts.csv", "rows-addenda/openjev-fp8-generate.cnb.idle.csv", "rows-addenda/openjev-fp8-generate.cnb.jsonl", "rows-addenda/openjev-fp8-generate.cnb.report.json", "rows-addenda/openjev-fp8-generate.cnb.watts.csv", "rows-addenda/openjev-fp8-readout-32k.along.idle.csv", "rows-addenda/openjev-fp8-readout-32k.along.jsonl", "rows-addenda/openjev-fp8-readout-32k.along.report.json", "rows-addenda/openjev-fp8-readout-32k.along.watts.csv", "rows-addenda/openjev-fp8-readout-32k.blong.idle.csv", "rows-addenda/openjev-fp8-readout-32k.blong.jsonl", "rows-addenda/openjev-fp8-readout-32k.blong.report.json", "rows-addenda/openjev-fp8-readout-32k.blong.watts.csv", "rows-addenda/openjev-fp8-readout.cnb.idle.csv", "rows-addenda/openjev-fp8-readout.cnb.jsonl", "rows-addenda/openjev-fp8-readout.cnb.report.json", "rows-addenda/openjev-fp8-readout.cnb.watts.csv", "rows-arm3/census.openjev-bf16-largecard.json", "rows-arm3/census.openjev-fp8-largecard.json", "rows-arm3/openjev-bf16-largecard.a.idle.csv", "rows-arm3/openjev-bf16-largecard.a.jsonl", "rows-arm3/openjev-bf16-largecard.a.report.json", "rows-arm3/openjev-bf16-largecard.a.watts.csv", "rows-arm3/openjev-bf16-largecard.b.idle.csv", "rows-arm3/openjev-bf16-largecard.b.jsonl", "rows-arm3/openjev-bf16-largecard.b.report.json", "rows-arm3/openjev-bf16-largecard.b.watts.csv", "rows-arm3/openjev-bf16-largecard.c.idle.csv", "rows-arm3/openjev-bf16-largecard.c.jsonl", "rows-arm3/openjev-bf16-largecard.c.report.json", "rows-arm3/openjev-bf16-largecard.c.watts.csv", "rows-arm3/openjev-fp8-largecard.a.idle.csv", "rows-arm3/openjev-fp8-largecard.a.jsonl", "rows-arm3/openjev-fp8-largecard.a.report.json", "rows-arm3/openjev-fp8-largecard.a.watts.csv", "rows-arm3/openjev-fp8-largecard.b.idle.csv", "rows-arm3/openjev-fp8-largecard.b.jsonl", "rows-arm3/openjev-fp8-largecard.b.report.json", "rows-arm3/openjev-fp8-largecard.b.watts.csv", "rows-arm3/openjev-fp8-largecard.c.idle.csv", "rows-arm3/openjev-fp8-largecard.c.jsonl", "rows-arm3/openjev-fp8-largecard.c.report.json", "rows-arm3/openjev-fp8-largecard.c.watts.csv", "rows-gaps-card/kev-4b-raw.kc.idle.csv", "rows-gaps-card/kev-4b-raw.kc.jsonl", "rows-gaps-card/kev-4b-raw.kc.report.json", "rows-gaps-card/kev-4b-raw.kc.watts.csv", "rows-gaps-card/kev-4b.ka.idle.csv", "rows-gaps-card/kev-4b.ka.jsonl", "rows-gaps-card/kev-4b.ka.report.json", "rows-gaps-card/kev-4b.ka.watts.csv", "rows-gaps-card/kev-4b.kc.idle.csv", "rows-gaps-card/kev-4b.kc.jsonl", "rows-gaps-card/kev-4b.kc.report.json", "rows-gaps-card/kev-4b.kc.watts.csv", "rows-gaps-card/kev-4b.kperm.idle.csv", "rows-gaps-card/kev-4b.kperm.jsonl", "rows-gaps-card/kev-4b.kperm.report.json", "rows-gaps-card/kev-4b.kperm.watts.csv", "rows-gaps-card/kev-9b-raw.kc.idle.csv", "rows-gaps-card/kev-9b-raw.kc.jsonl", "rows-gaps-card/kev-9b-raw.kc.report.json", "rows-gaps-card/kev-9b-raw.kc.watts.csv", "rows-gaps-card/kev-9b.ka.idle.csv", "rows-gaps-card/kev-9b.ka.jsonl", "rows-gaps-card/kev-9b.ka.report.json", "rows-gaps-card/kev-9b.ka.watts.csv", "rows-gaps-card/kev-9b.kc.idle.csv", "rows-gaps-card/kev-9b.kc.jsonl", "rows-gaps-card/kev-9b.kc.report.json", "rows-gaps-card/kev-9b.kc.watts.csv", "rows-gaps-card/kev-9b.kperm.idle.csv", "rows-gaps-card/kev-9b.kperm.jsonl", "rows-gaps-card/kev-9b.kperm.report.json", "rows-gaps-card/kev-9b.kperm.watts.csv", "rows-gaps-card/openjev-fp8-generate-mm.shot_kind.idle.csv", "rows-gaps-card/openjev-fp8-generate-mm.shot_kind.jsonl", "rows-gaps-card/openjev-fp8-generate-mm.shot_kind.report.json", "rows-gaps-card/openjev-fp8-generate-mm.shot_kind.watts.csv", "rows-gaps-card/openjev-fp8-generate-mm.shot_title.idle.csv", "rows-gaps-card/openjev-fp8-generate-mm.shot_title.jsonl", "rows-gaps-card/openjev-fp8-generate-mm.shot_title.report.json", "rows-gaps-card/openjev-fp8-generate-mm.shot_title.watts.csv", "rows-gaps-card/openjev-fp8-generate.creorder.idle.csv", "rows-gaps-card/openjev-fp8-generate.creorder.jsonl", "rows-gaps-card/openjev-fp8-generate.creorder.report.json", "rows-gaps-card/openjev-fp8-generate.creorder.watts.csv", "rows-gaps-card/openjev-fp8-readout-mm.shot_kind.idle.csv", "rows-gaps-card/openjev-fp8-readout-mm.shot_kind.jsonl", "rows-gaps-card/openjev-fp8-readout-mm.shot_kind.report.json", "rows-gaps-card/openjev-fp8-readout-mm.shot_kind.watts.csv", "rows-gaps-card/openjev-fp8-readout-mm.shot_title.idle.csv", "rows-gaps-card/openjev-fp8-readout-mm.shot_title.jsonl", "rows-gaps-card/openjev-fp8-readout-mm.shot_title.report.json", "rows-gaps-card/openjev-fp8-readout-mm.shot_title.watts.csv", "rows-gaps-card/openjev-fp8-readout.creorder.idle.csv", "rows-gaps-card/openjev-fp8-readout.creorder.jsonl", "rows-gaps-card/openjev-fp8-readout.creorder.report.json", "rows-gaps-card/openjev-fp8-readout.creorder.watts.csv", "rows-gaps/rerank.a.jsonl", "rows/base-generate.a.idle.csv", "rows/base-generate.a.jsonl", "rows/base-generate.a.report.json", "rows/base-generate.a.watts.csv", "rows/base-generate.b.idle.csv", "rows/base-generate.b.jsonl", "rows/base-generate.b.report.json", "rows/base-generate.b.watts.csv", "rows/base-generate.c.idle.csv", "rows/base-generate.c.jsonl", "rows/base-generate.c.report.json", "rows/base-generate.c.watts.csv", "rows/base-readout.a.idle.csv", "rows/base-readout.a.jsonl", "rows/base-readout.a.report.json", "rows/base-readout.a.watts.csv", "rows/base-readout.b.idle.csv", "rows/base-readout.b.jsonl", "rows/base-readout.b.report.json", "rows/base-readout.b.watts.csv", "rows/base-readout.c.idle.csv", "rows/base-readout.c.jsonl", "rows/base-readout.c.report.json", "rows/base-readout.c.watts.csv", "rows/census.gemma4-26b.json", "rows/census.hf.co_mradermacher_jevify-gemma4-26b-a4b-GGUF-Q4_K_M.json", "rows/census.openjev-fp8.json", "rows/jev-generate.a.idle.csv", "rows/jev-generate.a.jsonl", "rows/jev-generate.a.report.json", "rows/jev-generate.a.watts.csv", "rows/jev-generate.b.idle.csv", "rows/jev-generate.b.jsonl", "rows/jev-generate.b.report.json", "rows/jev-generate.b.watts.csv", "rows/jev-generate.c.idle.csv", "rows/jev-generate.c.jsonl", "rows/jev-generate.c.report.json", "rows/jev-generate.c.watts.csv", "rows/jev-readout.a.idle.csv", "rows/jev-readout.a.jsonl", "rows/jev-readout.a.report.json", "rows/jev-readout.a.watts.csv", "rows/jev-readout.b.idle.csv", "rows/jev-readout.b.jsonl", "rows/jev-readout.b.report.json", "rows/jev-readout.b.watts.csv", "rows/jev-readout.c.idle.csv", "rows/jev-readout.c.jsonl", "rows/jev-readout.c.report.json", "rows/jev-readout.c.watts.csv", "rows/openjev-fp8-generate.a.idle.csv", "rows/openjev-fp8-generate.a.jsonl", "rows/openjev-fp8-generate.a.report.json", "rows/openjev-fp8-generate.a.watts.csv", "rows/openjev-fp8-generate.b.idle.csv", "rows/openjev-fp8-generate.b.jsonl", "rows/openjev-fp8-generate.b.report.json", "rows/openjev-fp8-generate.b.watts.csv", "rows/openjev-fp8-generate.c.idle.csv", "rows/openjev-fp8-generate.c.jsonl", "rows/openjev-fp8-generate.c.report.json", "rows/openjev-fp8-generate.c.watts.csv", "rows/openjev-fp8-readout.a.idle.csv", "rows/openjev-fp8-readout.a.jsonl", "rows/openjev-fp8-readout.a.report.json", "rows/openjev-fp8-readout.a.watts.csv", "rows/openjev-fp8-readout.b.idle.csv", "rows/openjev-fp8-readout.b.jsonl", "rows/openjev-fp8-readout.b.report.json", "rows/openjev-fp8-readout.b.watts.csv", "rows/openjev-fp8-readout.c.idle.csv", "rows/openjev-fp8-readout.c.jsonl", "rows/openjev-fp8-readout.c.report.json", "rows/openjev-fp8-readout.c.watts.csv", "run.py", "run_addenda.py", "run_arm11.py", "run_arm8.py", "run_gaps_card.py", "run_kev.py", "tables.py", "tables_addenda.py", "tables_arm2.py", "tables_gaps.py", "tables_gaps_card.py" ] }, "seat_class": { "writer": "gemma-class", "judge": null, "writer_runtime": "vllm" }, "bench": { "bullets_written": 3, "bullets_kept": 1, "digests_written": 19, "digests_kept": 16, "chips_written": 8, "chips_kept": 8, "chips_grounded": 0, "chips_published": 0, "dropped_by": { "new_noun": 0, "figure": 1, "length": 0, "cite": 0, "judge": 0, "quote": 0, "directive": 0, "redaction": 4, "profanity": 0, "identifier": 0, "bare_figure": 0, "house_voice": 0 }, "new_noun_tokens_checked": [ "0.046", "0.085", "0.1228", "0.141", "0.2375", "0.350", "0.724", "0.8", "0.83", "0.9", "0.945", "000", "01Z", "1", "10", "100", "108", "12", "15", "151", "168", "17", "2.8", "20", "200×", "2026-09-21", "2026-09-22", "210", "25", "27", "27-billion-parameter", "290-token", "3090", "33", "34", "36", "4", "40", "40×", "43Z", "5.1", "5.7", "500", "598-megabyte", "6000", "620", "80", "80.4", "82", "84.0", "85.2", "85.4", "87", "9", "91.4", "93.5", "96.6", "97.2", "AI", "Astra", "Brier", "Darwin", "FREE", "Fable", "GPT-6", "Ibn", "Jev", "Kev", "Kev-9B", "LLM", "OpenJev", "OpenJev's", "PRO", "Qwen3.8-27B", "RTX", "Sagan", "Sina", "Socrates", "System", "UTC" ], "new_noun_tokens_withheld": 4, "figure_definition": "v3", "figure_boundary": "guarded", "drops": [ { "kind": "digest", "reason": "redaction", "item": null, "item_chars": 277, "withheld": "a string carrying a fenced literal", "detail": "" }, { "kind": "digest", "reason": "redaction", "item": null, "item_chars": 227, "withheld": "a string carrying a fenced literal", "detail": "" }, { "kind": "digest", "reason": "redaction", "item": null, "item_chars": 244, "withheld": "a string carrying a fenced literal", "detail": "" }, { "kind": "bullet", "reason": "redaction", "item": null, "item_chars": 75, "withheld": "a string carrying a fenced literal", "detail": "" }, { "kind": "bullet", "reason": "figure", "item": "the accuracy for the six-way task across six shuffles was a mean of 91.4 %", "detail": "terminal token not verbatim in the cited span" } ] }, "generated_utc": "2026-09-23T01:19:21Z" }