Updated
Glossary
Terms
590 definitions from the Terms block on each article. A word can appear in more than one story. The wording here follows the explainer when we have one, otherwise the longer note.
- __none__
- A required option in jym's Choice, alongside the real subcommands. If Jev picks it, the wrapper does not substitute a documented command.
- -27.0%
- Astra input tokens on the September 24 Extra High run, 19,198 to 14,011, six checks still passing. Medium is 19,204 to 14,013. The September 22 three-task run cut files from 28 to 4 and cut Astra input 1.2%.
- @cf/cloudflare/clef
- The Workers AI run id for Clef. clef-flash uses @cf/cloudflare/clef-flash. typesafe/jev on the same host is the Jev listing, a different card.
- $0.000096936
- Wes Bos's billed cost, October 14, 2026, 23:15 UTC, for one question on an 850 by 850 image. The post says that is about 10,000 images for a dollar.
- $0.00044
- The article's cost for one Jev Choice rerank query in the offline test. p50 latency on that test is 0.195 seconds.
- $0.0015
- Printed cost of the 40 Jev decisions on that table. The without-Jev cell is $0.005 to $0.015. The article's 150-to-250 times estimate is a separate Opus rate-card line.
- $0.0027
- Estimated model cost for the Jev QuickBooks run, from model list rates. The Opus 5.5 cell is $1.38. The page says this is model usage only.
- $0.04
- The September 30 post's price, per million input tokens. The product page we opened did not print this rate. The post also says sub-second latency and 250M free tokens.
- $0.042
- Per million input tokens, the rate the September 28 note uses for jev-1.13.0. Output tokens are free in that formula. The note says this is not invoice reconciliation.
- $0.042 per million input tokens
- The input price on the October 9 Command Line post and on the OpenRouter page opened October 10. Output tokens are free on both. The Azure Foundry pricing page opened that day did not show the name Decision-1.
- $0.05 and $0.50 per million tokens
- Input and thinking prices on the TypeLLM homepage and in the API reference, opened October 10, 2026. Typed answers are free. New accounts get $5. A field that thinks on the hosted API gets a 1,024 token budget unless the caller sets another, up to 4,096.
- $0.06
- Drex Max input price on the docs page, per million tokens. Output is unbilled. The October 19 post prints the same rate. The product table also prints $0.06 per 1,000 decisions.
- $0.10
- Clairic input price on the Workers AI card, per million tokens. The card lists no output charge.
- $0.1112
- Total cost printed for the 60-commit run. The post's $0.11 rounds this figure. The package-root rule moved the earlier total from $0.1115.
- $0.30
- One Jev pass over 19 benchmarks that otherwise use an API judge. The paper says that is 63 times less than those judges at list prices. A September 26 post prints $18.96 on the judge side. That figure is not in the tables we transcribed.
- $0.542
- Jev's cost per 1,000 calls on resource ranking, against GPT-OSS 20B at $0.817. p90 on that row is 511 ms against 5,445 ms. The page calls this the highest-volume decision.
- $2.94
- TechCrunch's figure for one monitoring job with Jev, against $372 with a frontier LLM, described as in theory. The demo page does not print either number.
- $40 million seed
- The round SiliconANGLE reported on September 16, led by DCVC. Techmeme says PitchBook records that seed at a $200 million valuation. Palazzolo's September 24 post says a later round, still being raised, is above $1 billion at a valuation above $10 billion.
- $55
- The demo page's estimate per million actions with Jev, against about $6,900 for a frontier LLM judge. Median on that page is 0.23 seconds. About 70 checks a second from one laptop.
- 0 flips
- Jeva's printed decision-flip rate on 600 Choice items under original, reversed, and seeded random candidate orders. Jev-1.13.0 on that test is 17.00%.
- 0 illegal
- fit.md's count for jobs whose options are listed up front. 489 MOBA decisions and 37 chess moves, and none of the answers was an illegal move. The page treats that as what happens when an illegal answer cannot be written down.
- 0.02
- The registered smallest effect of interest in DH-3. P2 minus N0 is +0.0030, under that line. P1 minus N0 is -0.0025, opposite the registered expectation.
- 0.09 to 0.22
- The probability of yes on questions about an API behavior the snippet does not state, after the question names the behavior. The page's examples are Array.sort with no comparator, splice to the end, a regex without the g flag, and a promise that leaves the try.
- 0.1.0
- The version string on the Hugging Face card samratduttaofficial/WaterSheep, checkpoint id watersheep-20260928-125452.
- 0.119
- Selection ratio on that 60-commit run, the share of test files picked. The post's 12% rounds this figure. A wider re-export variant reached recall 0.946 at selection 0.208 and was rejected.
- 0.146694928
- API-equivalent dollars for the harness arm on Extra High. The note computes (14011 times 10 plus 126 times 50 plus 6784 times 0.042) divided by 1e6. The uninstrumented arm is 0.201630. This is not a Codex invoice.
- 0.168
- Median expected calibration error. A null that knows the base rate is 0.074. The paper says thresholds fit on one benchmark do not transfer.
- 0.19 / 0.17 / 0.13
- Box confidence on this desk's three October 16, 2026 jev-1.14.0 every-vehicle calls, Date headers 08:19:44 GMT to 08:19:45 GMT, on a three-car PNG. Each answer was one rectangle.
- 0.19 s
- TypeSafe P50 on OpenRouter's typesafe/jev-1.14 page, opened October 16, 2026. The October 15 reading printed 0.15 s. Uptime 100.00 percent. GET /api/v1/models still omitted the id. This desk did not call OpenRouter.
- 0.2.1
- The live board edition in the September 28 file. 38 benchmarks. Arts and Human Taste weighs 10%. The other four areas share 90% by the square root of their benchmark count. Gold benchmarks weigh 1.2 inside an area. The file lists Jev 1.13.0 at 57.91. The kit's test locks that cell at 57.89.
- 0.25
- The tie band in index02.ranks. Two scores within 0.25 index points share a rank. AutoJev-27B at 50.94 is 0.73 below Jev at 51.67, outside that band. On the September 27 file, Decider chat Gemma-4-31B at 51.93 is 0.26 above Jev, also outside that band.
- 0.254
- Weighted mean this desk computed from jev-1.13.0 answers on DeepEval's Tool Faithfulness example, where the answer adds a light breeze and 40% humidity. Weights 2, 1, 1, 1. Date header 01:58:11 GMT on October 14, 2026.
- 0.3
- Opened October 8, 2026, the space header says v0.3, 115 models, Jev jev-1.13.0, updated 2026-10-07. Perplexity Decider v1.1 is still first at 62.8. Jev is a reference at 60.1. Opened October 7, 2026, the header said 112 models, updated 2026-10-06.
- 0.300
- The line next to the five ratings in the README screenshot: 0.760, 0.555, 0.468, 0.345, and 0.345. The screenshot labels the group as five skills added.
- 0.42
- Laya, English instructions, on the October 2 chart. The multilingual Laya row on the same chart is 0.21. Typed Laya is 1.46. These are log-AUC cells, not accuracy percents.
- 0.46
- Mean Noul on this desk's three October 15, 2026 jev-1.14.0 calls, Date headers 06:16:36 GMT to 06:16:40 GMT, asking whether that same PNG contained a red square. Scores 0.49, 0.47, 0.42. Matching jev-1.13.0 mean 0.94.
- 0.47
- On five identical jev-latest calls, the Choice label changed for 2 of 42 claims. Both flips stayed under confidence 0.47 on every run. Confidence changed on 28 of 42. The verdict file says no claim that reached 0.5 on any schema-only run changed label, across 420 calls under both model names.
- 0.5
- The README's bar. If Jev's best relevance estimate is below 0.5, the original MCP result is left unchanged.
- 0.50
- The abstract's micro-F1 on UNFAIR-ToS before thresholds are tuned on training data. After that tuning the same line is 0.75. Binary probabilities are described as ranking well and sitting poorly against a fixed 0.5 cutoff.
- 0.51
- Choice confidence on this desk's October 14, 2026 trap shortlist, Date header 02:35:12 GMT, jev-1.13.0. The pick was a closed API 504 issue at probability 0.55, with 0.43 on none.
- 0.56
- Mean red-square Noul on this desk's three October 15, 2026 jev-1.14.0 calls, Date headers 14:57:24 GMT to 14:57:25 GMT, on a 400 by 400 PNG that also held a blue square. Scores 0.55, 0.55, 0.57. Matching blue Noul 0.98. Matching jev-1.13.0 red mean 0.87.
- 0.6
- ASK_BELOW in agent_v1_permission_gate.py. At or above 0.6 the shell command runs. From 0.2 up to 0.6 the user is asked. Below 0.2 it is refused.
- 0.69
- The predicted-probability ceiling marked on Jev's calibration curves. The ensemble curves continue toward 1. The chart says the ensemble had no calibration step.
- 0.722
- winnow:e4b in the Ollaya README, against TypeSafe Jev at 0.738. The page also prints 89 ms for five questions on an RTX 4090. The item count for that pair is not defined on the page we read.
- 0.746
- Jeeves on the README's transfer overall, MMLU-Pro and buried state. Kev-9B is 0.579. The Kev-published Jev cell is 0.800.
- 0.80
- The sample deny band for destroys_uncommitted. At or above that probability the hook denies. The ask band on the same harm starts at 0.35. The evaluation note says these cuts are still starting guesses.
- 0.802
- Plumb v5 on the author's public hard tier, 89 of 111. The starting JevK5 v0.2 checkpoint, re-run on the same RTX 4080 Super, is 0.739, 82 of 111. McNemar p = 0.039.
- 0.815
- jevos accuracy on 2,000 yes/no questions from unseen policies, as printed in the README image alt. Jev on that line is 0.927. Laya is 0.489.
- 0.833
- Three-run mean accuracy on Yuuki Yamashita's 60 Kyoto-ben tickets, October 15, 2026, model jev-1.14.0. Matching jev-1.13.0 mean 0.817. Majority vote matched on all 60. McNemar p = 1.000.
- 0.842
- Custom-policy accuracy at a 0.5 cutoff on the PolicyLM card. Policy edits followed is 0.528. The H100 median on that row is 22 ms. The table has no Jev column. The card says every set except OR-Bench was also used in development.
- 0.846
- v5 on House's held-out check: teacher questions from domains training never sees, held-out public splits, and a hand-written hard set. Models are chosen on this check. v4 was 0.844.
- 0.85
- Default deny threshold. Ask is 0.55. If authorized is at least 0.8, some axes soften one step. secret_exposure, violates_constraint, and unresolved_choice do not soften.
- 0.865
- Jeeves on the README's public JevBench hard tier, 111 items, with thinking. The Kev-published Jev cell on that row is 0.730. Kev-8B is 0.451. The sealed judge tier is not in this row.
- 0.886
- Median AUROC of a generic Noul over 31 benchmarks that have a Noul form. The interval is [0.821, 0.952]. Split-half targeted wording on 38 benchmarks is 0.911.
- 0.889
- Jeeves on the README's test overall, out-of-domain and held-out, item-weighted, with thinking. Kev-9B is 0.822 and the Kev-published Jev cell is 0.857. The 2,962-item split without thinking, 0.804, is a different line.
- 0.893 / 0.290
- Datadog's 40-run mean is_risky on a viewer who cannot edit, then on the same call with a counterfeit is_risky definition in profile_description. LangChain blocks at 0.5.
- 0.9 points
- The frozen cascade's accuracy margin over GPT-6 on 1,610 held-out pairs, at 41% of GPT-6's fee. The cascade escalates 31% of those pairs.
- 0.90
- The README's act-on-it line. Among decisions the system takes at confidence 0.90 or higher, error is 2.4% for Eikos-4B and Eikos-27B, 5.9% for Jev, and 37.3% for Laya, on 7,140 items across 6 suites.
- 0.901
- Weighted mean on the same questions when the answer is only 18°C and sunny. The Choice dropped because nothing_missing held 0.89 of the mass. Three applicable questions.
- 0.91
- Mean Box confidence on this desk's three October 15, 2026 jev-1.14.0 calls, Date headers 06:16:36 GMT to 06:16:40 GMT. Mean edges left 0.40, right 0.79, top 0.16, bottom 0.71. An empty field the same minute returned confidence 0.05.
- 0.926
- Mean recall on the September 28 re-run of 60 commits, jgrep 0.5.0 plus the package-root rule. The post's 90%+ rounds this figure. Code-only recall on the same run is 0.449.
- 0.93
- Box confidence on this desk's October 15, 2026 jev-1.14.0 call, Date header 01:22:15 GMT. Edges were left 0.4, right 0.8, top 0.16, bottom 0.7, on a red square drawn at 0.40, 0.80, 0.20, 0.70.
- 0.94
- Mean Box confidence on this desk's three October 15, 2026 jev-1.14.0 calls, Date headers 14:57:24 GMT to 14:57:25 GMT, on a blue square drawn at 0.08 to 0.28. Mean edges left 0.07, right 0.30, top 0.14, bottom 0.29. A Box on jev-1.13.0 the same minute returned HTTP 400.
- 0.984
- jevmod's AUROC on the sexual labels in OpenAI's moderation eval, 1,680 messages, remeasured September 23, 2026.
- 0.99
- Mean Noul on this desk's three October 16, 2026 jev-1.14.0 calls, Date headers 00:28:47 GMT to 00:28:48 GMT, asking whether a Checkout PNG contained a green Continue button. Scores 0.98, 0.99, 0.99. Matching jev-1.13.0 Noul 1.0. An empty field the same minute scored 0.02.
- 1,081 seconds
- The Jev arm's elapsed time in results.json, 1,081.05 seconds, which is the post's 18.0 minutes. The other arm is 1,480.50 seconds, the post's 24.7 minutes. The file records characters and call counts. It does not record the post's $0.155 and $0.089.
- 1,369
- Decisions in a measurement window on version 0.9.1, September 21 to 27. A blind second labeling covered 455 of them and found zero permissive misses. The README says 0.11.0 is the response to the ask-rate on that log.
- 1,372 ms
- DeepSeek V4.1 Flash on the page's later per-model chart. The two DeepSeek task rows above it print 1,466 ms and 1,456 ms. The page does not say how the 1,372 ms rollup was weighted.
- 1,536 image tokens
- Liquid's blog prices a 1024 by 1024 image at 1,536 input tokens, 1.5 tokens per 32 by 32 pixel patch. Images are billed at the same input rate as text.
- 1,995 of 2,000
- Emails the 400M model marked as phishing. True positives 1,000, false positives 995, true negatives 5, false negatives 0. AUROC 0.34821.
- 1.0
- The ai_decide function version in the SQL page and the REST reference. It is the default. The options map can set it with version.
- 1.2 percentage points
- How far rewording stays from the abstract's re-run baseline on jev-1.13.0. Fields outside the schema, the abstract says, never reach the model.
- 1.3% of perturbations
- The average rate at which eight perturbations of one request changed the decision, in the October 9 post. Paraphrased option descriptions, and reversed or shuffled options, had zero flips.
- 1.30
- Mean frustration Score on this desk's three October 15, 2026 Chinese payouts tickets, Date headers 06:16:21 GMT to 06:16:24 GMT, model jev-1.14.0. Scores 1.27, 1.32, 1.32. Mean confidence 0.54. Matching jev-1.13.0 mean 1.54 at 0.31.
- 1.37
- Frustration Score on this desk's October 15, 2026 Chinese payouts ticket, Date header 03:03:53 GMT, model jev-1.14.0, English questions. Confidence 0.44. The matching jev-1.13.0 Score was 1.54 at confidence 0.31.
- 1.7 points
- How far JEV sits behind GPT-6 on the paper's pooled 4,850 public judgments, eighth of fifteen LLM judges. The fee on that line is $0.044 per 1,000 judgments.
- 1.8% to 3.5%
- DH-4 validated attack success on fresh calls. Success means at least 3 target selections in 5 calls. Budget 0 is 1.8%. Budget 24 is 3.5%. The intervals are [0.0, 6.3] and [0.0, 10.0].
- 1/34
- Abe's cost ratio on those 1,774 posts: Jev about $0.055 against AI_FILTER about $1.87, at published list prices. The token-unit ratio in that article is about 1/77.
- 100 English judgments
- The count in the September 23 post. Laya's English checkpoint is scored 92 correct and 8 wrong. Jev is scored 100 correct. The author says the set is not a full benchmark.
- 100M free input tokens
- The homepage and the October 12 post offer 100 million free input tokens to the first 10,000 sign-ups. The blog calls the product Public Preview, first come first served, with a waitlist after capacity.
- 107 of 108
- Aviz Maeir's October 15, 2026 count of correct boxes when Jev 1.14 confidence was 0.7 or above. Luna printed 0.9 to 1.0 on 27 confident misses.
- 11435
- The port in the Ollaya README. POST /v1/systemone and /v1/decisions. TYPESAFE_BASE_URL is set to http://localhost:11435.
- 12,000
- Auto mode's source-byte threshold. The 75-byte main.py makes zero Jev calls. The 27,278-byte gardening file is what the selection pass is for.
- 12.1%
- The abstract's rate at which one unverified opinion, appended to the state, flips a jev-1.13.0 decision. The same sentence says this is statistically tied with the strongest injected command at 10.1%.
- 125
- The decide command's exit code when the Jev call fails. A choice returns the option index. A boolean returns 0 or 1 at the threshold.
- 127.3 seconds
- Jev's total time on the one QuickBooks bill in the write-up. Opus 5.5 on the same harness and the same 123 decisions took 978.9 seconds. Both checked 30 of 30 rows during entry.
- 14 decisions per second
- The rate in Inception's September 30 post and on OpenRouter's Mercury 2.5 page. The post does not print a JevBench score beside it.
- 15 of 20
- Pick-and-place scenes completed in the card's October 3 section. The card says every cube that was grasped ended in the tray, 15 of 15, and that each miss was a grasp 3.0 to 3.7 cm off the cube.
- 15.7 seconds
- The footnote's p95 for Solar Decide. Jev's p95 on that line is 0.40 seconds. The headline speed, 0.14 seconds against 0.30 seconds, is the median.
- 150 milliseconds
- The New Stack extract's attribution: OpenAI says the Decisions model returns results in 150 milliseconds, and that GPT-6 Luna would take 1.6 seconds. We did not open the full article. The post itself does not print this pair.
- 150,317
- Evaluation requests in the card's self-run. The card says every request was answered and none were truncated.
- 16 requests
- The hello-world smoke test, 3 turns. The interactive repeat of that smoke records 20,426 input tokens. The three-task ladder is a separate table, and none of those tasks pass.
- 18 of 21
- ImajevBench unknown items the 4B abstains on after phase 3. The previous release abstained on 14 of 21. It also abstains on 9 of 258 answerable items, up from 3.
- 19 of 20
- Frozen set 6 on toolgate 0.15.0. Ten matched pairs. All ten pairs separated. Zero permissive errors. The highest risk score on an inert side was 0.32.
- 19.3 s / 61.7 s
- Cloudflare's damaged-claim cascade, Clairic then Clef then GLM-5.3, against GLM-5.3 at every step. Cost on that table is $6,032 against $11,740 per million claims.
- 199/231
- Drex 1.5 on Nace's September 30 chart, 86.2% of 231 public JevBench items. Jev 1.13.0 is 201/231, 87.0%, on the same card. Hard tier is 82/111 for both.
- 2 cents
- His figure for the day of testing. The thread does not print token counts or a rate card.
- 2,238
- Input tokens on this desk's October 15, 2026 jev-1.14.0 call, Date header 11:04:54 GMT, one Noul, an 850 by 850 PNG in state. List-price arithmetic $0.000094. Two instruction references of that file used 4,209 tokens.
- 2,383
- Questions per checkpoint on which the README says fp32 exports match PyTorch, 100% of them.
- 2.3
- AI BENCHY score for Jev 1.14 default, opened October 15, 2026. Jev 1.13 default on that page is 2.6. Tests correct 4 of 12 against 6 of 12. The suite uses adapters to supply choices.
- 20 of 20
- AskUp facts Sung Kim says Solar Decide kept for long-term memory, against a keyword rule at 14 of 20. Casual chat stored by mistake is 0. The post does not run Jev on those 20.
- 20 of 21
- Lead qualification on jevbench.app. 10 human yeses and 11 human nos, recorded September 27, 2026. The printed miss is Jev yes at 0.51 where the human label is no.
- 2026-09-19-public-v5
- The frozen recorded run on RZ's dashboard. One hundred synthetic cases, five passes, 1,500 decisions. Earlier v2 through v4 directories are rejected by corpus hash.
- 205 of 231
- The same blog's later paragraph, the decision adapter on JevBench's 231 public items. It is not the composite 72.5.
- 228 of 231
- Correct tasks in TypeLLM's September 23, 2026 thinking summary for RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead on 231 public JevBench tasks. The no-thinking file is 195 of 231. The homepage prints those rates as 98.70% and 84.42%. The hosted name typellm-latest is not this checkpoint.
- 25.8%
- Total cost including a Jev list-price estimate, $5.6564 against $7.6221. Sol alone in that file is $4.1780. An older Sol-only comparison, Jev excluded, is $5.44 against $7.62, printed as 28.6%.
- 26-candidate cap
- Sarashina Choice rejects 27 or more options with HTTP 422. The same README says LFM and ModernBERT still accept 255. Jev's documented cap is 255.
- 262k
- 262,144 tokens. Diogo Almeida's October 14, 2026 tweet says context lengths should be up to 262k. TypeSafe's recap calls that 4 times as much as the default 64k per-request cap. 262,144 / 64,000 is 4.096.
- 29 of 30
- SosyalDijital's October 15, 2026 score for both Jev 1.13 and Jev 1.14 on 30 synthetic Turkish buyer, seller, and other messages. Both models marked the same buyer as a seller. Recorded through OpenRouter as typesafe/jev-1.13-20260917 and typesafe/jev-1.14-20261014.
- 31/38
- The README's reproduce line for the task-disjoint DeepSWE holdout: 31 of 38 tasks, printed as 81.6%. The head is fine-tuned. Opus 5 wrote the candidate solutions. Jev, as the verifier on the same setup, is described as scoring below pass@1.
- 32,768 tokens
- The context window on the Foundry catalog opened October 10, 2026. OpenRouter's header that day printed 33K. OpenRouter's October 9 post said 32K.
- 33 of 41
- Deadlines the rebuilt pipeline found, in the same post as the 75 of 78. The thread does not say what the other 8 pages looked like.
- 37 of 40
- Correct matches, right issue and right status, on Akshay's 40-request test. Both the Jev arm and the Opus-only arm print this count. On October 13, 09:21 UTC, he wrote that 40 is too small to establish how reliably confidence routes errors to review.
- 37 of 41
- About how many candidates, in an average query, received the same score from Jev Choice. The paired control is 4,855 captured queries. Recall@6 moved from 40.2% to 44.3% when production order broke those ties.
- 38%
- The README's aggregate drop in agent time for the default hook against Claude Code with no hook. Agent cost on the same sentence is 26% lower. TypeSafe's own token bill is not in that 26%.
- 384K tokens
- The Decisions docs window, 393,216 tokens. The October 12 post calls that 12 times Jev. TypeSafe's models page lists 32k for state plus the longest question and 64k for the whole request.
- 39.1%
- The page's reduction from $0.76199 to $0.46369 per 1,000 calls, across the five decisions it rolled up. The same box prints p90 latency from 4,752 ms to 508 ms, 89.3%.
- 4,458
- Input tokens on this desk's October 16, 2026 jev-1.14.0 call, Date header 00:28:50 GMT, four Nouls and four Boxes on one four-color PNG in state. Four separate POSTs summed to 6,612 input tokens.
- 4.5 times quicker
- Microsoft's October 9 speed line against Quyet-1.0-Large. The same post says 35 times quicker than GPT-6 Sol, and that P50 latency is about 35 times faster than GPT-6 Sol. It does not print the milliseconds.
- 41.63
- Vela-2.0-9B on Decision Index 0.2.1, 38 benchmarks, with noul_calibration on. The 9B card says the default with that option off is 41.09. The evaluation file prints the Lux base at 46.23 and does not print a Jev row.
- 422
- The HTTP status jevos returns for Choice and Score. Noul is the question type the server answers. Criteria text is accepted and not read.
- 43.5%
- Indirect prompt-injection attack success for the tool selector. The self-gated bar is 21.6%. Gemini-3-Pro on that panel is 55.7%.
- 446 of 492
- Jev 1.13 Armor-default agreements on OpenRouter's synthetic coding set seeded from harness logs. Clef is 435 of 492. Sonnet 5 is 434 of 492. GPT-6 Luna decisions is 407 of 492.
- 48.57
- d1-3B on the card's self-run of Decision Index 0.2.1, official scorer, not a leaderboard submission. d1-omni-600M on the omni card is 15.95. Winnow-12B on the 3B table is 50.02.
- 488 coefficients
- The paper's count of fitted weights, 100 in the Task 1 ridge models, 332 in the Task 2 rerankers, 32 in the pair-level valence-arousal maps, and 24 in category fusion. They are fit on CPU. The Jev weights are not updated.
- 5 of 12
- Coverage on Jev4Mellea's October 15, 2026 twelve-scene capture, model jev-1.14.0, reject 0.10 / accept 0.90. Accepted 0 of 6 positives. Uncertain 7 of 12. Started 11:45:44 UTC.
- 5/5 against 1/5
- The comments row, 196 KB, five runs each. JevScout was correct on all five. The no-hook path was correct on one. The other three printed rows are 5/5 against 5/5.
- 50.25%
- The same checkpoint on 2,000 PhishNChips emails, threshold 0.5, no fine-tune. It flagged 1,995 of 2,000. The published Jev reference on that set is 62.6%. Jev was not called again for this run.
- 5172/5172
- Valid JEV 1.13 judgments in Table 1. GPT-6 Astra is 5171/5172. The base set is a 642-item pilot plus 4,530 held-out items.
- 517d0f0
- The commit on the latest ladder row, after call-nesting and bash fixes. Guessing-game, file-io-script, and multi-file-package all still fail. Search width 3 fails the first two and does not run the third.
- 52.31
- Drex DLM on the card's Decision Index 0.2 table, against the kit's Jev cell of 51.67. The hosted Drex 1.5 row on edition 0.2.1 is 58.28.
- 55.8%
- Benign task success for the self-gated Jev row. The tool selector without that second check is 60.9%. Opus-4.7 on the same panel is 90.1%.
- 554 decisions
- The offline accuracy set, no time pressure. Qwen3.5-4B Q8_0 is 77.8%, Gemma 4 E4B Q6_K is 81.6%, Qwen3.5-9B Q5_K_M is 82.7%. The README says the questions were tuned on the 4B.
- 56.5
- Decision-2.0-Vega-27B on the authors' Decision Index 0.2.1 reproduction, opened October 5, 2026. The card does not print a Jev row. The September 28 board's Jev cell is 57.91.
- 56.6 ms
- The card's median per request, one MI355X, bf16, one request at a time. The post's text says 56.8 ms. The card says the board remeasures latency on its own RTX PRO 6000.
- 58.9
- d1's index on Liquid's September 29 chart, an internal reproduction of Decision Index 0.2.1. Jev 1.13 is 57.9 on the same row. The difference printed is +1.0.
- 59 times
- The largest cost bar on the chart, metric pick against the frontier LLM in use today. Escalate-to-human is 49 times. Root-cause tagging is about 45 times, estimated from per-token pricing. PII is 20 times.
- 59.26
- Matilda on Maincode's self-run of the Decision Index 0.2.1 kit, posted September 30. The card's Jev (hosted) cell is 57.91. The card says the model is being submitted, so this is not yet an independent board row.
- 6 of 8
- The opening card's count of quality checks at parity or better. The chart's two losses are messy metric questions, down 16.6 points, and 3-level ticket tags, down 3.6 points.
- 60.0
- Overall intelligence for jev 1.14.0 on TypeSafe's internal chart posted October 14, 2026. jev 1.13.0 on that chart is 58.9. The image is watermarked antibench.com and says Please don't trust them.
- 61.9
- imajev-4b capability on the JevBench v1.6.1 open-weights list. Opened October 8, 2026, revision v1.7.21, the rank line says capability rank 31 and official composite 47. Intelligence 34.6, Calibration 89.2, estimated $0.017 per 1,000 decisions, p50 0.29 s. Opened October 7, revision v1.7.12, the same score was capability rank 17 and official composite 26.
- 62.48
- The card's adaptive balanced skill from its own Decision Index 0.2.1 scoring, against the September 28 board's Jev cell of 57.91. Raw 70.66, breadth 62.00. Not a row in that board file. Adaptive mode misses the kit's latency cap on Knowledge and Reasoning.
- 63.8
- Sieve-31B Full score on Decision Index edition 0.3.2, board data of 2026-10-14, on the Hugging Face card opened October 20, 2026. The authors' public-suite run on that card is 62.52.
- 63.88
- StartLux-Decision-27B on the authors' Decision Index 0.2.1 run, opened October 4, 2026. The README says the run is not on the September 28 board, where Jev 1.13 is 57.91.
- 64.81
- GLiDE's balanced score on Fastino's own run of the Decision Index 0.2.1 scorer. The blog says the model leads 31 of 38 benchmarks and is not on the public leaderboard, because 0.2.1 submissions are paused.
- 65.84
- Plumb-4B on the JevBench v1.4.2.1 composite, changelog date September 27. v1.4.2.2 then placed Imajev-4B first at 67.37. The capability view we opened lists Plumb at official rank 2.
- 67.37
- Imajev-4B on the v1.4.2.2 composite, changelog date September 27. Plumb-4B on v1.4.2.1 is 65.84. Jev remains 63.29 on that board. The capability view we opened September 28 lists Imajev at 66.3 and official rank 1, and Jev at 64.7 and official rank 4. v1.5 does not reuse these cells.
- 67.59
- Drex Max on Nace's product page, Decision Index 0.3.2, public benchmarks only, opened October 20, 2026. Jev 1.14 on that bar is 60.70. 140,178 scored requests, 37 benchmarks.
- 69.12
- StartLux-Decision v1.1 27B Instant Full score on Decision Index 0.3.2, README release notes dated 2026-10-16, opened October 20, 2026. First of 174 entries. Public 72.88, same-skill private 64.82, new-domain private 61.41.
- 7,210-row panel
- The chart's fixed set, dated September 2026. Eleven named benchmarks plus an overall row. Jev 84.51%, Qwen3.8-27B 74.76%, pplx-decider-v1-27b 85.71%.
- 70.1%
- Direct misuse attack success for Jev 1.13 as the tool selector, on the September 27 chart. Self-gating, a second Jev check after the tool call is filled, is 36.1% on the same chart.
- 70.4
- imajev-4b on JevBench v1.5.6, the page opened October 4. Official rank 10, roster addendum A2. The same point estimate was rank 9 on the v1.5.4 page opened September 30, and would-place sixth on the v1.5.0 page opened September 29.
- 71.24
- mammoth-1 Full score on Decision Index 0.4, first of 208, on the Hugging Face card and the technical report. The live space opened October 23 is edition 0.4.1 and prints 68.4.
- 71.5
- Jev 1.13.0's composite on the JevBench API board, release v1.6.1. Opened October 8, 2026, revision v1.7.21, the page ranks 26 hosted systems and this row is fourth. Capability on that card is still 77.1, rank 2. Opened October 7, revision v1.7.12, the same score was rank 3 of 23. 72.1 is the v1.5 composite.
- 718 of 720
- Orderings, in the September 22 fair-die post, where TypeLLM + Qwen gave the highest probability to the first option. Jev gave it to the label one in 720 of 720 orderings, and to the first option in 120 of 720. That post is not the JevBench file.
- 73.7
- Cygnet on JevBench v1.5.0 official score A, the page we opened September 29. Axes 71.1, 87.0, 91.0, 56.4. Winnow-12B Q8 is 73.2. Jev 1.13.0 is 72.1. The page calls Cygnet and Winnow a statistical tie. 63.29 is the older v1.4.2.2 cell, not this score.
- 75 of 78
- Posts Deji had labeled no that the rebuilt rules hid. 5 of 5 posts he labeled yes were shown. The thread's earlier count of real yeses in the 100 is 7.
- 75 queries
- The subtitle of Rox's chart. We did not open a methods note, a query list, or a repository. The post's 20 times faster and 10 times cheaper are not drawn on the chart.
- 76.4
- imajev-4b on Image JevBench v0.1.5, the page opened October 4. Official rank 2 of 50. The compare table prints 76.39. The author's September 28 screenshot of v0.1.3 prints 76.39 as rank 1 of 49.
- 79.96%
- skundu42/kev on 17,476 held-out rows from its own prepared split. Jev 1.13 on the same rows, through OpenRouter, is 73.54%. A 100-row pilot went the other way: 80.00% and 81.00%.
- 8 of 10
- Official SWE-bench solves for both the jevgrep arm and the saved no-Jev baseline, on ten tuned Python tasks. matplotlib-26466 and pylint-4604 stay unresolved.
- 8,192
- The context length the README prints for jevos, in tokens. The same note says Laya is 512 and does not state a number for Jev.
- 8.1×
- Median per-entry speedup on 40 golden entries where Glean's existing router actually transferred to an expert. The wider routing set is 751 entries. The accuracy on those 751 is not in the text we could copy.
- 812
- Typed questions in the abstract, across 66 scenarios. The abstract also counts 9,744 single-edit variants. We did not read the 19 tables behind those counts.
- 82.90% / 3.88%
- Datadog's Noul has_prompt_injection recall on 4,643 labeled attacks, raw then with a counterfeit policy prepended. True positives 3,849 to 180. Threshold 0.5.
- 82%
- Aviz Maeir's October 15, 2026 score for Jev 1.14 Box on 180 scenes against OpenAI Responses API plus strict JSON. gpt-6-luna 84 percent. gpt-6.1-sol 95 percent. Median 0.21 s. $0.15 per 1k. This desk did not rerun the 180.
- 83.9%
- imajev-4b on ImajevBench after the September 26 phase-3 adapter. The previous release was 82.4%. The set is 279 questions, including 21 whose labeled answer is unknown.
- 84.07
- Six-group mean on the card's September 27 table. The hosted Jev cell on that table is 83.85. The JevBench column inside the mean is not Florian's official board.
- 87.0%
- Solar Decide on the chart's public JevBench 231. Jev 1.13 on the same card is 86.1%. The card does not print how many items those percents are.
- 90.5%
- jev-latest accuracy on the authors' 1,248-request general set. fsdecide on that file is 60.5%. The doc says the set has 240 cases in 12 domains. It is not Florian's JevBench.
- 91.3%
- fsdecide on the 80-item review suite in the October 5, 2026 report. jev-latest on the same items is 56.3%. The interval is 83.0% to 95.7% against 45.3% to 66.6%.
- 91% / 9%
- HiddenLayer's TypeSafe Jev API print on a simulated checkout after a Confirm Price ad, among clickable buttons. Confirm Price 91 percent, Place order 9 percent. The jev-ultrafast harness is in that write-up.
- 96%
- Jev-Score at 200 chunks on Rox's September 28 chart, mean over 75 queries, relevance kept. Production, labelled GPT-5 Mini, is 84% at the same point.
- Abort-on-finish
- The Opus 5 column is stopped when Jev completes the wire, so the slower model is not billed for the remaining headlines.
- about 1,000 tokens
- The README's figure for the 20-skill storefront demo: initial system prompt with the picker, against the same prompt carrying the whole catalog.
- about 10,000 tokens
- The post's figure for a 250-plus skill repository at work. The README does not describe that repository.
- Accessibility tree
- The structural snapshot of a webpage exposing element roles, labels, and states. These workflows supply this tree to Jev as state rather than transmitting screenshot images.
- Accurate when confident
- In this write-up, accuracy where a pick-one confidence is 0.9 or higher, or a yes/no probability is at or below 0.1 or at or above 0.9, plus the share of items in that band.
- Action budget
- The attempts argument on test(). It is a cap on UI actions executed, not on scenario restarts. Goal evaluation also runs before the first action and after the last allowed action.
- Action space
- The set of valid moves available for the active piece, recomputed on each turn similarly to how browser agents rebuild interactive DOM elements.
- Action table
- The compact state Jev sees: eligible paths, road boundaries, nearby traffic, signals, stop memory, and destination guidance. Shared values are sent once.
- African Typed Decisions
- olaverse/african-typed-decisions. The translated comparison is 3,010 cases and 15,050 decisions in eight languages. KL, Brier, and ECE on the card are for that translated set.
- AG News training sample
- MotherDuck's published prompt_jev table uses 100,000 rows drawn with a reservoir sample, seed 43, from the AG News training parquet. It is not the dataset's held-out test split.
- Agent Arena net
- Arena's October 7 figure for typesafe/jev-router, +3.6% net improvement at a median $0.155 per task, on more than 4,700 agent sessions. The post says that would rank below DeepSeek V4.1 Flash.
- agents per laptop
- The gist's count of how many browser processes fit in free RAM. It launches N copies, holds them open, and samples peak resident memory. The default page is a Hostelworld listing for San Francisco.
- agreement gate
- HydroJEV accepts a Jev verdict without calling the reviewer only when the verdict is benign, normal transient or sensor fault, and the hand-written rule tree R1 returns the same class. Every other window goes to the reviewer.
- AI Gateway
- A cloud provider's managed proxy for AI APIs handling credentials, billing, and logs, such as those provided by Vercel, Cloudflare, and Netlify.
- ai_decide
- Databricks AI Function in Beta on the September 30, 2026 blog. SQL is ai_decide(state, questions). REST is POST /api/2.0/ai-functions/ai-decide. A question is a noul, a choice, or a score.
- AI_FILTER
- Snowflake Cortex function that classifies a prompt as TRUE or FALSE. Abe compared it with Jev Noul on 1,774 of his 2025 posts.
- anchored scale
- A jsort JSON file of anchor texts and their scores. Later texts are compared only with those anchors, so a new line can be placed without sorting the original file again.
- answer slot
- Rizzo Flow maps each candidate to one uppercase letter that must be a single token. Twenty-six letters means at most 26 options. Abstention uses a slot too. The README contrasts that with Jev at 255 and SemIf at 16.
- Artificial Analysis Decision Index
- A four-eval composite from Artificial Analysis, weighted equally across AA-Omniscience-Decision, HLE-MultiChoice, When2Call, and WildGuardTest. It is not Hugging Face Decision Index 0.2.1 or 0.3.
- ASO Workbench
- Mustafa Kendigüzel's Mac app for App Store keyword ranks. A bundled CreateML model estimates popularity where Apple does not publish a number. Jev, through OpenRouter, answers brand, demand, and language-fit questions used as features.
- assets
- Top-level map on a System One request. Each image is listed once, keyed by a reference string that is written into state, instructions, or criteria. Docs opened October 15, 2026.
- authored144
- SemIf's committed fixture, scored with SemIf's own evaluator. Rizzo Flow Q8_0 on CUDA is 0.812 balanced accuracy, 0.793 on the held-out half. SemIf's published Q8 cells are 0.819 and 0.811. Half of the rows are the dev split the prompt was chosen on.
- AutoModeMiddleware
- An experimental wrap_tool_call hook where Jev evaluates tool call safety, returning an error ToolMessage when an action exceeds risk thresholds.
- Awesome list
- A curated catalog of projects. Two independent lists appeared during launch week; project counts on each reflect point-in-time snapshots.
- Backbone
- The pretrained language-model checkpoint a decision model keeps. The decision head reads its hidden states or its option logits.
- Bag of decisions
- GPT-5 writes yes/no questions for one query. Jev answers each as a Noul. The reranker adds 10 times the sum of those probabilities to the candidate's retrieval score.
- Balanced accuracy
- Unweighted mean of per-class recall over vulnerable, safe, and insufficient context. The published v5 run puts Jev at 65.3%, Terra at 75.1%, and Opus at 91.2%.
- banking feed
- In this post, rows from a product that categorizes bank transactions. The 612 rows are described as a sanitized, harder subset of an evaluation set of about 33,000. The categories are not listed.
- basal-1.5
- Remek Kinas's Apache-2.0 decision family for Polish and English. Three sizes, 1.5B, 4.5B, and 11B, on a Bielik base. The server is basal-serve, POST /v1/systemone.
- basal-omni-1.0-ultra
- Remek Kinas's Apache-2.0 31B checkpoint on Gemma 4, served by basal-serve at POST /v1/systemone. Opened October 19, 2026, JevBench ranks it first of 184 open systems on Capability at 82.5. The card prints Werdykt v1 at 0.879. Text and images. No audio.
- battery
- For sixteen accusation-type sentences, the same judgement is also asked eight ways, three of them reversed. A needs_battery sentence does not publish its single reading alone. 187 of 335 names under those sentences are not quoted as numbers.
- Blink parameter counts
- The card text says 4B active parameters and 17.5 GB on disk. The Hub listing's safetensors total is 14,386,941,232 parameters, and the page chrome says 14B.
- Blink readout
- The card says one question per prompt, thinking off, and the answer is a softmax over option letters at the first generated token. The shipped temperature is 1.3.
- Blink v0.3
- Pixilab's decision checkpoint on Gemma 4 26B-A4B. The card's license field is Apache-2.0. Its self-run of Decision Index 0.2.1 is 57.48, against the board's Jev cell of 57.91.
- boolean, vegaml 0.5.0
- The yes/no answer type in vegaml 0.5.0. The README says the name replaced a coined word and does not print the old word. The loader remaps older checkpoint keys.
- Borderline-weighted pairs
- Lemkin's 200-pair slice that over-represents close cases. He reports 59.5% agreement with Sonnet there, against 67% on 200 random pairs.
- bot challenge
- A page that asks the client to prove it is human. The script treats a challenge as a failed comparison, because both engines can be measured against a stub. The comment names booking.com as a case that would print a fictional 36 times.
- Box
- System One question type on Jev 1.14. type is box. The answer is four edges, top, bottom, left, and right, each from 0 to 1, plus confidence. Docs opened October 15, 2026. A Box on jev-1.13.0 returned HTTP 400.
- Bradley-Terry here
- Hunter's MM algorithm on soft wins, prior 0.1, log-strengths centered on a geometric mean of 1. A gap of 1 means the higher item wins about 73 percent of a head-to-head.
- brain
- The Jev channel used for routing. The README lists TypeSafe, OpenRouter as typesafe/jev-1.13, OpenCode Zen, and Vercel AI Gateway. With none configured, jevonian/auto returns an error instead of picking a model.
- BranchOne
- Necati Demir's October 2026 open System One checkpoint. Qwen3.5-0.8B-Base plus LoRA and a shared scoring head return one probability per supplied candidate. Apache-2.0. Hugging Face ndemirdemir/BranchOne. His A100 Colab median on ten short text questions is 0.305 s.
- budget_ms
- Span-02's thinking deadline in milliseconds. auto thinks when the top answer is below 0.8. on always thinks. off never thinks. The blog's no-limit Decision Index cell is 64.1.
- Calibrated decision
- TypeSafe's phrase, used in the TechCrunch piece, for a typed answer with a probability rather than generated text.
- Calibration
- The degree to which a model's reported probability corresponds to empirical label accuracy, which TypeSafe targets via RLCD.
- Capability Score
- JevBench's mean of Intelligence and Calibration, ranked among systems inside twice Jev's cost and median latency caps. Opened October 19, those caps are about $0.0646 per 1,000 decisions and 1.23 s.
- CC BY-NC 4.0
- The license on the StartLux-Decision weights. The code in the repository is Apache-2.0. Commercial use of the weights needs a separate license from StartLux Labs.
- celeris-1-decision
- Celeris's hosted decision model, posted October 8, 2026. Model id celeris-1-decision. POST https://inference.celeris.ai/celeris-1-decision/v1/systemone, and POST /v1/decisions on the same host. Input $0.04 per million tokens, output free. The blog prints 80.0% on 3,210 jev-bench items at a 67 ms median. Opened October 16, Artificial Analysis Decision Index lists it at 52.
- chess score
- On the Eikos Chess page, wins plus half the draws. The September 24 screenshot scores 388 games as 53.9% for Eikos-27B.
- Choice
- Selecting one option from a user-defined set (cardinality up to 255), returning a probability distribution and confidence score.
- CHR
- Conditional hit rate at a Jev confidence of 0.7 or higher. The paper prints 0.2% at budget 24 and 0.2% at the baseline checkpoint.
- ClaimBench
- Vals AI's 1,960-claim 10-K excerpt test. October 22 post: d1-ultra is the only decision model to get all 1,960 right. The page, opened October 23, prints 99.9% alone at a 1% error budget, 0.00% wrong. 1,960 is not Artificial Analysis 65.
- Clairic
- Cloudflare's System One extraction model, preview on Workers AI as @cf/cloudflare/clairic, posted October 19, 2026. A request is a state plus typed fields. The reply is a value of the declared type, or null, with confidence and p_null.
- Classic set, OpenBles
- deepset/prompt-injections test split. The README prints 55 of 60 injections caught and 2 of 56 benign prompts flagged. Jev on that split is 33 of 60, with no false flags.
- classifier.dev fast tier
- A hosted classification API whose own page, and JevBench's note, say the fast tier is Jev. It matches Jev on Intelligence and leads the ranking on a cheaper published plan and a faster measured p50 from the bench's server.
- Clef
- Cloudflare's 27B decision model, Apache 2.0, hosted as @cf/cloudflare/clef. The model card says it is post-trained from Qwen3.8-27B.
- Clef serving times, October 9
- Median and p95 on the hosted Clef model after the October 9 server change. About 800 tokens moves from 262 / 438 ms to 152 / 351 ms. About 3,400 tokens moves from 616 / 777 ms to 305 / 531 ms. About 16,000 tokens moves from 2,721 / 3,250 ms to 1,635 / 1,805 ms. This is not the October 1 median of 209.3 ms.
- Clef-flash
- The 9B model in the same release, hosted as @cf/cloudflare/clef-flash. The model card says it is post-trained from Qwen3.5-9B.
- Clef-flash price, October 9
- The October 9 blog lists hosted Clef-flash at $0.038 per million input tokens and a 24k context window. The October 1 card listed $0.09 and 65,536 tokens. The blog says the Hugging Face weights were trained for 256k and were not replaced.
- Clef-omni
- Cloudflare's October 9, 2026 decision model, hosted as @cf/cloudflare/clef-omni at $0.15 per million input tokens. Apache 2.0 weights from Qwen3-Omni-30B-A3B-Instruct. The blog's BANKING77 cell is macro-F1 94.8.
- coach facts
- In jev-rubiks, the plain-English description code sends after a hand move: beginner stage, pieces placed, whether the move undid progress, and how long the coach has been quiet. Jev answers whether to speak. It does not return the next algorithm.
- Compaction
- Reducing an agent's context history so subsequent prompts fit model limits. Traditional approaches rely on summarization prompts; this method scores entries and removes irrelevant ones.
- Confidence band
- A slice of answers above a probability cutoff. On this set he reports zero errors in Jev's most confident 85.5%, and that nearly half of answers below 70% were wrong.
- confidence statistic
- TypeSafe metadata on a Jev Choice, keyed by question id. The form router reads it from providerMetadata.typesafe.confidence. It runs from 0 when probability is spread across destinations to 1 when it sits on one destination, and it is not the probability of the selected option.
- Context filter
- AstroHan's loop: Jev answers one boolean per ~2k-character tool chunk (state is task, intent, action, output) and the harness keeps the chunk when p > 0.5.
- Continue IoU
- Mean IoU 0.93 on this desk's three October 16, 2026 jev-1.14.0 Continue-button boxes, Date headers 00:28:47 GMT to 00:28:48 GMT. Mean edges left 0.55, right 0.91, top 0.71, bottom 0.88, confidence 0.91. Drawn 0.55 / 0.90 / 0.70 / 0.88.
- continuous F1
- The paper's extraction score. An exact span match contributes a weight that falls as the valence-arousal error grows, and the reported figure is 100 times the macro average over eight corpora.
- Contrastive data curation
- Write two examples that differ in one fact so the correct answer flips. The rest of the context and the question stay the same.
- Control plane
- In this PR, a classifier lane that labels each GPT Live utterance before the main model runs. Config values are off, agent, or classify.
- counterfeit policy
- Datadog's payload inside untrusted state. It names the answers to return, quotes the original message as a benign exception, and rewrites attack-category definitions. Trusted Noul and Choice instructions stay as the developer wrote them.
- d1-ultra
- Liquid AI's hosted decision model, posted October 15, 2026. Model id d1-ultra. POST https://api.liquid.ai/decisions/v1/systemone. Input $0.042 per million tokens, output free. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57.
- d1:free
- The model id on Liquid's decision-model docs. The page we read does not print a paid model id or a paid rate.
- d3
- vLLM Semantic Router's 26.09B Decision 3.0 checkpoint, Apache-2.0, from Qwen3.8-27B. Opened October 26, 2026, Decision Index 0.4.1 Full score is 67.15. The October 16 card reading of edition 0.3.2 was 66.3. The free API aliases vllm-sr/d3 to d3-turbo.
- d3 vision row
- The d3 card's vision-board 0.3.1 table prints 71.6, public 74.2, private 68.9. The same table prints JEV-27B-VL at 69.6, public 72.8, private 66.4. Perception Test on that card is 77.4 of 19,140 questions.
- d3-turbo
- vLLM Semantic Router's upgraded 26.09B Decision 3.0 checkpoint, Apache-2.0. Opened October 26, 2026, Decision Index 0.4.1 Full score is 67.31. Vision full on edition 0.4 is 72.46. The free API aliases vllm-sr/d3 to this model.
- d3-ultra
- vLLM Semantic Router's 26.09B Decision 3.0 checkpoint, Apache-2.0, from Qwen3.8-27B. Opened October 26, 2026, Decision Index 0.4.1 Full score is 70.18, second of 229. The card's private SR-Text mean is 87.2.
- Decider 1
- meraGPT's hosted model sd-1. The September 22 blog says it answers noul, choice, and score questions in one call, and that a typesafe-sdk client can point TYPESAFE_BASE_URL at https://meragpt.com.
- decider-4b v2
- Mapika's open row at the top of the v1.4.2 harmonic mean, 64.13. v1.4.2.2 keeps that cell and ranks Imajev-4B and Plumb-4B above it. Intelligence is 49.4 against Jev at 53.1. The cost cell is an estimate from a Qwen3.5-4B list price, not a GPU bill.
- Decision 2.0
- Six Apache-2.0 checkpoints from vLLM Semantic Router, posted October 3, 2026. The cards name Kai, Eos, Sol, Nox, Lux, and Vega. A call returns a probability for every listed option and does not write a reply.
- Decision 3.0 family
- Six Apache-2.0 cards opened October 11, 2026. Their own 0.3.1 index cells are d3 64.7, d3-flash 59.0, d3-mini 54.2, d3-nano 37.2, d3-lite 29.0, and d3-edge 18.0.
- Decision API is not enabled
- The HTTP 403 sentence from POST https://api.openai.com/v1/decisions on October 2, 2026, with this desk's key, and again from an empty body on October 4. everruns pull 3924 recorded the same sentence on September 30. On October 6 a call with model gpt-6-luna returned HTTP 200, and an empty object returned HTTP 400, missing model.
- Decision head
- The small stack trained to score a fixed set of options. Clef, Kev, Laya, and Unsloth each publish their own version of it.
- Decision Index
- A chance-corrected average of 40 public benchmarks in five equal areas. Zero is random guessing and 100 is perfect. Unanswered requests count as wrong. Edition 0.2 is dated September 24, 2026. The kit's test locks Jev at 51.67.
- Decision LCR
- Subquadratic's conversion of Artificial Analysis AA-LCR into 118 multiple-choice decisions over 72K-115K token document sets. The blog prints Snap 74.0 one-shot against Jev 48.0 chunked, and says the scores cannot be compared with published AA-LCR results.
- Decision Model Factory
- Yuma Ichikawa's recipe pipeline for specialized decision models on Gemma 4 backbones. arXiv 2610.34956. Table 1 prints DF-JA-Large at 67.63 on S1MB Japanese Task Avg after specialization.
- Decision rate
- How often the sim asks Jev for a new path. The README caps this at 4 Hz in traffic and about 1.5 Hz on an empty road. Single-answer questions are resolved locally.
- decision_weight
- The multiplier on the summed Noul probabilities. Both published configs set it to 10.
- decision-v7
- Author 10-source typed decision suite used in the MiniCPM5-2B-Jev README. The project prints 86.41% accuracy and ECE 0.0307 on that suite.
- Decisions API
- Venice's endpoint POST /api/v1/decisions, which accepts state and typed questions rather than a chat transcript. OpenRouter uses the same name for POST /api/alpha/decisions.
- degrees@f518286e
- The scheme hash for this corpus. Five Score questions share one state. Each criterion has five levels, 0 through 4, worth 5 points on a 0 to 20 scale. Five criteria sum to the 0 to 100 score the tables print. Rewording a criterion changes the hash.
- Development set
- Together's 1,000-item main set and 300-item policy-transfer set. The card says the mixture informed training, so the 88% and 100% are not an untouched benchmark.
- DF-Long
- The factory's long-context specialist on Gemma 4 26B-A4B, about 4B active parameters per token. Table 2 prints S1MB Japanese 63.57 and JQaRA nDCG@10 90.1.
- DiffusionGemma as Jev
- A one-step read of google/diffusiongemma-26B-A4B-it. The canvas is seeded with the answer template, and the label slots stay noisy. mmastrac/djev, Apache 2.0, is the example server in front of vLLM pull 57250.
- diode
- Jevsus's main test. A statement is offered as a forced true/false Choice, both option orders. The README also publishes a native yes/no probability, the same Choice with a decline option, and the forced pick under a perspective prefix. The four are not merged.
- Direct logit readout
- One forward pass over declared option tokens, with no sampled answer. SemIf's 3090 note: 1.023 s for 21 questions versus 5.332 s for a compact JSON array from the same weights.
- direct mode
- A jev-gateway outcome. Jev is confident about the tool, and every argument is an enum, a boolean, or a constant, so the gateway writes the tool call and does not call the language model.
- Direct question
- One configured Noul, with true and false criteria, asked of each of the top 100 products. The blog labels this column jev_reranker. The repo calls it bag_of_decisions_direct.
- djev
- Maisa's hosted decision API on DiffusionGemma, and the name Matt Mastracci used for DiffusionGemma-as-Jev. JevBench v1.2.8 lists the hosted row at 74.3.
- djev-run
- Daniel Lee's Cloud Run recipe, GitHub taeold/djev-run. It serves mmastrac/djev-spark on an NVIDIA RTX PRO 6000 and exposes POST /v1/systemone. The repository page lists no license.
- DocJev
- Jerry Liu's Apache 2.0 library for classifying and splitting PDF, DOCX, and PPTX files. LiteParse reads pages locally; hosted Jev returns the category or the page-range boundaries.
- does_pass
- The binary evaluator in jev-as-a-judge. Jev matched the human oracle on all 500 repeated decisions in this run.
- Dr. GRPO
- In this repository, the advantage is the rollout reward minus the mean reward of the other rollouts for the same question. The code does not divide by the group's spread.
- Drex DLM
- The card and repository title for the weights Nace posted on October 6, 2026. The post calls the release Drex 1.1. The card does not print 1.1. Backbone NVIDIA Efficient-DLM-8B. POST /v1/systemone. Weights CC BY-NC 4.0. Nace code MIT.
- drex-max
- Nace's pinned hosted id for Drex Max. POST https://console.nace.ai/v1/systemone. Docs say drex-latest does not point at it, and a request that omits model is answered by drex-v2.0.
- Dry run
- Operating against live market data without private transaction keys, simulating order fills when trade prints cross the resting quote.
- dual encoder
- One encoder reads the state and another reads the candidate action. Both are a frozen Qwen3-8B plus a 20M projection head. The answer is a softmax over dot products. Action embeddings can stay cached while the state changes.
- duckdb-jev
- Prasanth J's Apache 2.0 native C++ DuckDB extension. SQL functions call TypeSafe's Jev over HTTP, with batching, streaming, and an optional connection cache.
- E4c
- The fresh sealed set from look-alike family C, 200 windows. The agreement gate settled 76 of them. Accept precision there is 0.579, and 21 of the 76 accepted windows were attacks.
- Early warning
- Taylor's description of using Jev as a first pass that is cheap enough to run even when a later human or larger model still reviews the result.
- elephant-1
- Base Labs 9B sibling, MIT, from Qwen3.5-9B. Card Decision Index 0.4 62.65. Live 0.4.1 Full 60.3. Artificial Analysis replica 65.2. Hosted model decision-models/baseten-elephant-9b-latest.
- EmbeddingGemma 2
- Google DeepMind's October 6, 2026 multimodal embedding model, 740M parameters under Apache 2.0, modular text/vision/audio encoders, 768-d vectors, 8K context. Hugging Face id google/embeddinggemma-2.
- evaluate
- Vercel AI SDK's experimental method for Jev, passing state and questions to receive typed answers rather than text completions.
- evidence unit
- The slice asked about in one request: a function, a file outline, a pair of copies, or a test. Tree-sitter finds it on the machine. The upload is that slice, bounded by upload_allow and upload_deny.
- Expected Calibration Error
- A score for how far stated probabilities sit from observed accuracy. Lower is better. Poly reports 0.081 for ModernBERT and 0.105 for Jev on this Banking77 run.
- Fail open
- Every error path returns the original input as steer, so a classifier outage does not drop the utterance.
- fail_open
- LiteLLM's default when Jev is unavailable: send the original request to the generator without compaction.
- fallbackReason
- The form router's label for why Luna decided. null when Jev cleared 0.95. low-confidence, missing-confidence, or jev-error otherwise. A failed Luna call does not return a reason or a destination.
- False admission
- A negative case the model treats as acceptable. Lemkin reports 47 of 148 for Jev (32%) and 20 for Sonnet (13.5%) on this SaaStr Connect sample.
- flat page search
- One Choice whose options are the pages, each described by its own text. The README says the request has to fit in 32k tokens and a Choice stops at 255 options.
- four scene types
- Feature descriptions, negation, mention versus the intended target, and conversational padding. The post gives one English example of each. The item list is the reply's report image.
- frozen template
- The abstract's setup for jev-1.13.0. One prompt template per dataset, held fixed, zero-shot, on the full evaluation split. The open reference models see the same requests and are scored by exact next-token probabilities over the options.
- fsdecide
- ForkScape's 149M ModernBERT-base encoder, release fsdecide-base-v11e4s21. The comparison doc says it was trained on about 34,000 labelled decisions from five app questions, and that the 450 bench items were kept out of training. The local server speaks POST /v1/systemone.
- fx auto mode
- A Vercel safety classifier Sharma says previously ran on gpt-5.6-luna. The September 16 post is the source for the 5 to 18 times figure.
- generally available
- OpenAI Developers, October 22, 2026, 19:07 UTC. The guide, opened October 23, prices uncached gpt-6-luna input at $0.10 per million tokens and cached input at the standard Luna cache-read rate. Output is uncharged. Date headers 02:23:12 GMT and 02:23:13 GMT October 23, two POSTs from this desk returned HTTP 200.
- Geometric mean
- The v1.2 formula. The four axis scores are multiplied and the fourth root is taken. A weak axis pulls the total down; Luna's cost score of 28.2 is why a 96.8 Intelligence row ranks 12th at 66.0. v1.4.1 replaces this with an equal-weight harmonic mean and a sealed set.
- GEV-26B-Decide
- AutoTrust's October 3, 2026 checkpoint. About 4B active parameters on google/gemma-4-26B-A4B-it. Previously released as JEV-Gemma4-26B-A4B. Adapter, head, and calibration are Apache-2.0. The base stays under the Gemma 4 license. Not affiliated with TypeSafe.
- GLiDE
- Fastino's decision model, posted October 1, 2026, with v1.2 on October 22. Opened October 26, Decision Index 0.4.1 Full score is 71.5, rank 1 of 229. Docs price input at $0.04 per million tokens. POST https://api.fastino.ai/v1/systemone, model fastino/GLiDE. Weights on Hugging Face under the Fastino Community License.
- GLiFormer
- Knowledgator's encoder with task heads at inference time. jeff loads knowledgator/gliformer-large-v1.
- gravity 0.2 s
- The Tetris setting on the Jev side of the snapshot. Opus 5.5 played at 0.4 seconds. Piece counts at the snapshot are 282 and 71.
- H2O-Lightning-4B v1.1
- H2O.ai's October 7, 2026 blog prints this checkpoint at JevBench composite 72.5 against Jev 1.13.0 at 71.5, on board revision v1.7.17. The blog says Apache 2.0, from Qwen3.5-4B.
- Hails 2026
- D. Hails, The CIA was 'Probably' Right, opened September 30, 2026. A Prolific survey of 99 English-speaking respondents with an undergraduate education or higher, 25 phrases, medians rounded to the nearest 5. Probable is 70 in the Hails column, and 70 in the NATO 1970s column and the Zonination 2015 column (n=46). The essay has no Jev column.
- hard 18
- Real sentences from one published article, paired with the passage they cite. Nine are wrong, six are right, and three cite a passage that does not cover them. One claim is 5.6 points. The other 24 claims are controls built from the same audits. Headline comparisons use the 18.
- Hard tier, 109 of 111
- The thinking summary's hard tier. The no-thinking file has 76 of 111. Easy in those files is 48 of 48 without thinking and 47 of 48 with thinking. Original is 71 of 72 and then 72 of 72.
- hard-deny
- An Armor rule that blocks a matching call. An allow rule and user consent do not clear it. Default match threshold 0.6.
- head_max_len
- Laya's option-prompt token budget. At default 256 on the multilingual checkpoint, 77 Banking77 labels get about 3 to 4 tokens each, which the README says is why accuracy falls to 0.425.
- Header denylist
- semgate.WithHeaderDenylist. Without it, Authorization and Cookie are sent to TypeSafe with the rest of the request.
- held-out sources
- The second split in the 12,000-row table. Accuracy 0.8366, ECE 0.0944, Spearman 0.976. The in-split test row is 0.9323 and ECE 0.0079. The README does not print a Jev score on either split.
- Helium
- A browser reported to cause account registration failures during the initial onboarding period.
- Hierarchical context
- A proposed structure organizing execution history into a nested task tree to allow logarithmic search instead of sequentially reading complete logs.
- high_risk floor
- In the AgentOS Jev strategy, a Noul at or above 0.7 forces the turn to tier c3 in code, so a short request to delete production data does not depend on the route Choice reading the words literally.
- hint mode
- The jev-gateway path for Claude Code. A short suggestion is added to the request because a forced tool_choice would break extended thinking or the prompt cache. The model can ignore it.
- HiringCafe Spearman
- Hamed Nilforoshan's October 7, 2026 thread. Query-to-job Spearman is 0.74 for Jev, 0.71 for OpenAI, and 0.63 for Gemini 3.1 FL. Resume-to-job is 0.71, 0.67, and 0.61. The thread does not print the pair count.
- Hit@1
- Share of labelled queries whose correct document is ranked first. On McLaren's 41-query TypeSafe-docs set, jevsearch is 83% and the keyword pass alone is 41%.
- hotdog-27B
- A rank-16 LoRA on dense Qwen3.8-27B. It reads the logits for A (false) and B (true) at one token position and returns a probability. Apache-2.0 adapter, Tinker format, selected at temperature 1.0.
- hunch
- Matt Mastracci's October 27, 2026 harness that serves EmbeddingGemma2 as a /v1/systemone endpoint on CPU, with published clocks of 20 ms on text and 100 ms on images, and a student that learns from a teacher such as Cloudflare Clef in real time. Open-source launch was announced for later that week. This desk has no public repo or hosted call yet.
- Hybrid override
- Postgres plans first. If Jev's estimate is confident, that plan can be replaced. The thread reports this avoided the 10x regression from blind overrides.
- Independent verifier
- A check of the app's actual state (file locations and hashes, Calculator display) that the decision policy cannot read.
- Injection Range
- semgate-example, a Cloud Run service with six simulated injection sinks. With a TypeSafe key the guard answers 403; without a key every payload reaches the handler. No real side effects.
- input battery
- The JevGuardrailAdvisor check on the user's request. A block returns before the chat model is called. The output battery checks the model's answer and does not retry.
- Intelligence-per-cost
- Almeida's term for the trade-off between computational cost and decision quality, prioritized over raw demonstration latency.
- Jaggedness
- TypeSafe's word for uneven failure modes on jev-1.13. The docs page lists nine of them and says many will be fixed in later versions.
- JamC-QA
- The third column on t.toda's chart. Jev is printed at 56.0% and GPT-5.6 Luna at 76.5%. The post suggests this column is where Japan-specific detail shows up. Item count is not on the image.
- jeff
- A self-hosted server that speaks TypeSafe's System One API and answers with Knowledgator's GLiFormer instead of Jev.
- jev branch
- The jev branch of kishida/llama.cpp. It builds a modified llama-server that serves POST /v1/systemone and GET /jev.
- Jev RAG
- A name used for a typed Jev stage after retrieval and before a language model writes. Avi Chawla's October 7 post and aifabrice/jev-rag use it. TypeSafe's pages for the same stage are titled Re-ranking and Classifying RAG passages.
- Jev Router
- OpenRouter's typesafe/jev-router, posted September 25. The announcement says Jev scores the conversation, and the router picks a model and an effort. The thread's four-benchmark count is 237 of 423 against Auto Router's 130 of 423. Görn's September 28 log of the same model id names gpt-6-luna, deepseek-v4.1-flash, and kimi-k3 as the models that answered.
- jev_ask
- A Snowflake function that sends several named questions about one row in a single TypeSafe call, so the row is billed once.
- jev_plan
- A tool the agent calls with three to five hypotheses and a read-only test for each. Jev ranks those tests against a fresh namespace snapshot and does not run them.
- jev_stream
- A DuckDB table function in duckdb-jev. It takes a subquery of row id, evidence, and questions, packs HTTP batches across chunks, and returns answers keyed by the supplied ids.
- jev_submit
- A review gate before diagnosis or mitigation is sent to the grader. Every required question had to reach probability 0.70 or the agent had to gather new evidence.
- JEV_THRESHOLD 0.3
- The default cut in docs/jev.md. The page says it was not calibrated on the USDA queries, and that the Jev hit rates for this release were not measured.
- jev-1.14.0
- TypeSafe's October 14, 2026 Jev release. On October 15, 2026, Date header 01:22:13 GMT, a POST with model jev-preview returned this id. The models page prices it at $0.042 per million input tokens, output free.
- JEV-27B-VL
- AutoTrust's Apache-2.0 adapter on Qwen3.8-27B. System 1 is POST /v1/decide: yes/no, a choice of 2 to 256 options, or a score from 0 to 5, over text or images. System 2 is the unmodified Qwen checkpoint writing text.
- jev-bench 3,210
- Celeris's October 8 table uses Praveenrajus/jev-bench, 22 datasets, 3,210 items, harness 0cfd276. It is not Florian's JevBench composite on benchmarkheaven.com.
- jev-compaction
- LiteLLM guardrail that asks Jev whether each older tool result still matters for the latest user question, then replaces results below relevance_threshold (default 0.2) with a short notice.
- jev-latest
- TypeSafe's alias for the current stable model release, currently routing to jev-1.13.0.
- jev-preview
- Alias that, opened October 15, 2026, points at jev-1.14.0. GET /v1/models listed it with release_date 2026-09-10. The blog says jev-latest will move to 1.14.0 in a week.
- jev-ultrafast
- Browser Use's open harness that turns a page into a numbered list of controls and asks a System One model which to click. HiddenLayer ran the checkout through that harness.
- jev.batch_size
- How many rows realZachi/pg-jev packs into one System One request. Version 0.2.0 defaults to 20. The README's ground-truth check was 100% at 1 to 20 rows and 92% to 98% at 40.
- jev()
- In realZachi/pg-jev, a boolean PostgreSQL function. It is true when the Noul probability for a plain-language condition clears a threshold. The default threshold is 0.5.
- Jeva
- Qi Li and Xinchao Wang's System One architecture and the Jeva-4B checkpoint, posted October 14, 2026. Apache-2.0, from Qwen3.5-4B. The paper prints 0 flips on 600 Choice items and BANKING77 at K=16 91.62 against Jev-1.13.0 at 82.50.
- JevArena
- The score printed on the Decision 2.0 cards for a frozen prompt set. The footnote says a missing or invalid answer counts as an error. The cards do not print how many prompts that set holds.
- JevBench public subset
- The 231 public tasks in JevBench, out of 534. Original 72, easy 48, hard 111. The other 303 tasks are private or judge-only, so a score on this slice is not the full benchmark composite.
- JevBench Score
- Benchmark Heaven's composite for typed-decision systems: Intelligence, Calibration, Speed, and Cost at 25% each, combined with a geometric mean. v1.2.2 reports Jev 1.13.0 at 75.3.
- JevBench v1.4
- The edition the post cites. On this desk the v1.4.2.2 harmonic mean has Jev at 63.29. The live page on October 1 is v1.5.4, and a search of that text found no Mercury Decide.
- jevbench.app
- Yusuke's playground of recorded Jev runs. It is not fstandhartinger/jevbench and not the Benchmark Heaven ranking.
- jevctl
- npm package for Nasrallah-AL/jev-cli. The command is jev. It wraps TypeSafe's Jev as verify, screen, classify, extract, find, rerank, match, route, ask, compact, and batch.
- jevel
- A folder with one JEVEL.md. Frontmatter names the model, the required state fields, and the questions. jevelry reads that file. The body is for people.
- JevEval
- DeepEval's custom metric that sends bounded Noul, Score, and Choice questions to Jev in one request and returns a weighted mean in [0, 1]. Apache-2.0. This desk's invented-breeze case scored 0.254 on jev-1.13.0.
- Jevflake
- Luke Kranz's dbt package (and Terraform module) that creates Snowflake SQL functions for Jev's noul, choice, and score questions. Default model jev-1.13.0.
- Jevis
- A Dart package that runs on Flutter integration_test, sends the current screen and a registered action list to Jev, and executes the chosen widget action.
- JevJudge
- A Spring AI TypeSafe builder. Each criterion is a Noul or a Score plus the threshold it must clear. Confidence under the default floor of 0.5 is reported as INCONCLUSIVE rather than FAILED.
- jevmod
- Omar Hernandez's MIT package, version 0.2.1. One Noul per moderation category. Adult text flags at 0.80.
- jevonian/auto
- The virtual model that asks Jev once for the route and the thinking depth. A named phase, a real model id, or routing.mode set to off does not make that call.
- Jevons paradox
- The economic observation that increases in resource efficiency can expand total demand. The product is named around this economic hypothesis.
- jevsearch
- Kyle McLaren's MIT shadcn registry block for site search. A lexical pass answers first; Jev re-ranks the top candidates in one request.
- JGLUE test row
- The README's shared-API table. LFM: JNLI 17%, JCommonsenseQA 69%. Sarashina Q4: 34% and 86%. ModernBERT-Ja 310M, trained on the JGLUE train split: 93% and 92%. The page says this is not a fair ranking of model ability.
- Join order benchmark
- JOB, a standard set of join-heavy queries over IMDB-style data, used here as the reported test. The thread does not list which query subset or Postgres version.
- jsort
- Khaled Eltokhy's MIT command, installed as jev-sort. It asks Jev which of two texts ranks higher on a plain-English comparative, fits a Bradley-Terry scale, and prints the file in that order with a standard error.
- Judgment
- One typed answer to one question about one document. 37 documents times 21 questions produced the 777 count.
- Judgments model
- Jevflake's recommended dbt table: one stored answer per key and question, resent only when the row, the questions, or jevflake_model change.
- jym
- The command from syumai/jevyoumean. It wraps a CLI, and on an unknown subcommand asks Jev which documented command the token resembles, then can run that correction.
- keep-threshold
- jev compact's default 0.5. A tool call or its result is kept when Jev's probability is at least this. LiteLLM's similar guardrail defaults to 0.2.
- kettei-fat
- Sqwish Labs' hosted decision model, model id kettei-fat. POST https://api.sqwish.ai/v1/decide. Docs price input at $0.06 per million tokens, output free. Opened October 17, 2026, JevBench API board v1.7.46 ranks it first of 39 at composite 77.8.
- Kev
- Jared Palmer's Apache 2.0 family of Jev-like decision models. Rank-16 LoRA plus a pointer head on Qwen bases. Serves POST /v1/systemone.
- language Choice
- Four-option Choice this desk added to the payouts tickets: english, chinese, spanish, japanese. On October 15, 2026, Date headers 03:03:53 GMT to 03:03:54 GMT, jev-1.14.0 returned 1.0 on the matching label for each of the four states. Date headers 06:16:21 GMT to 06:16:35 GMT, a later sweep also returned 1.0 on those four, and on French, German, and Korean in a seven-option Choice.
- Late
- A status indicating the model response missed the block deadline, resulting in a hold action without updating quotes.
- Laya
- Convai Innovations' Apache 2.0 typed-decision engine. Three checkpoints plus a Router. JevBench v1.2.2 reports 70.1 for the Laya row against Jev 75.3.
- Legal landing
- A valid board placement calculated by the game engine. The model chooses exclusively among these options.
- letter logits
- Eikos reads the probability of each option from the letter tokens, in one forward pass over the shared state. More than 26 options go through a tournament. The served route is POST /v1/systemone.
- lfm1.0
- The license_name on both cards. The d1-3B LICENSE file is LFM Open License v1.0. Section 5 conditions commercial use on annual revenue under $10,000,000.
- limited preview
- OpenAI Developers' wording on September 29. Access was limited to selected API customers for testing. The follow-up said a broad release was planned in the coming days. On October 6 the Developers post said public beta for all developers. The DevDay recap, opened that day, still used the September wording. On October 22 the Developers post said generally available.
- list_fallback
- A field on the jev-router pipeline stage. The guide says the value models_ignored means the include list matched no pool model, so the router used the whole pool. excluded_models still applies.
- Listwise Choice
- One Choice whose options are the candidate passages. Hindsight 0.10.1 uses that shape. The probabilities sum to 1 inside the call, so the top score is a rank in that pool.
- LiteParse
- The default local parser in DocJev. It extracts complete page text on the machine. Decision timing in the published pilots excludes this step.
- Literal reading
- Answering the words in the instruction rather than the intent behind them. The docs say to put boundary cases in the criteria, or split the judgment into two questions.
- live shadow
- The speed chart's footnote for the 4 times escalation bar. Thousands of production conversations, run in parallel with the current model. Offline escalation on the same chart is 1.9 times.
- llama.cpp v0.6.0
- The October 5, 2026 release of ggml-org/llama.cpp. llama-server answers POST /v1/systemone for the October 2 decision GGUFs, for Bespoke-Nimble-9B-v3, and for Clef. The server README at that tag names OpenJev and Clef as the image models.
- llamajev
- The MIT command in NakliTechie/llamacpp-jev. It launches or attaches to an unmodified llama-server and serves POST /v1/systemone by reading single-token label logprobs.
- LLM fallback
- A small language model used exclusively when the chosen operation requires text input.
- LLM-as-a-judge
- A generative model scores another system's trace. LangChain compared that pattern with Jev on five frozen weather-agent runs.
- Local perception
- ScreenCaptureKit, OmniParser CoreML, Apple Vision OCR, and accessibility labels run on the Mac. Pixels and coordinates are not sent to Jev.
- LocalJev
- GitHub Next's local HTTP server that accepts TypeSafe's System One request shape and answers with a model running behind an OpenAI-compatible Chat Completions API.
- Macro accuracy
- Unweighted mean of the 13 subset accuracies in the public suite. Nimble's macro is 74.8%. The micro, 75.9%, counts each of the 3,880 records once. Ollama's later mean on those records is 75.7% for Nimble, with the Jev cell left at 76.0%.
- macro-F1
- The unweighted mean of the per-class F1 scores on the four causes. The paper's primary comparisons use a class-stratified paired bootstrap and Holm correction inside each round.
- mammoth-1
- Base Labs 27B decision checkpoint, MIT, from Qwen3.5-27B. Card Decision Index 0.4 Full 71.24. Live space 0.4.1 Full 68.4 rank 3, opened October 23. Hosted POST https://inference.baseten.co/v1/systemone, model decision-models/baseten-mammoth-27b-latest.
- MCP
- Model Context Protocol. typesafe-mcp and jev-mcp expose Choice, Score, and Noul as tools callable by Claude Code or similar clients.
- meriyah
- A DeepSWE task whose unfiltered tool output was 1.46 million characters. Unfiltered 0/49 tests; filtered 49/49.
- MiniCPM5-2B-Jev
- Community System One checkpoint on OpenBMB's MiniCPM5-2B, posted by OpenBMB on October 28, 2026. Project prints JevBench 78.79%, first among open models at or under 2B, and decision-v7 86.41% with ECE 0.0307. Apache-2.0. Weights ytbai/MiniCPM5-2B-Jev.
- Mizan 4B
- xmanii/Mizan-4B, Apache-2.0, a Clef-style decision head on Qwen3.5-4B. The card's mean on 3,000 real Persian records is 78.7, against a Jev cell of 83.7.
- Mizan Q4
- The card's Q4_K_M GGUF, listed at 2.5 GB. The same seven-task mean is 78.3. The card says that file loses 0.4 points against the 78.7.
- Mizan text limit
- The card says the vision tower from Qwen3.5 is still in the weights and was not trained for decisions. The October 11 post says image input is planned for a later update.
- MMAI Gym
- The training set named on the chart. The footer says 12,563 molecules. The post does not link a download or a paper.
- MMAU-mini
- A multiple-choice audio understanding set. Qwen-Audio-Jev prints 78.70%. ToddlerJev prints 68.70% on 1,000 items. Those two cells are not one run.
- Mock model
- The repository's default heuristic that simulates trading signals locally without requiring an API key.
- Model completion
- A Noul on the current screen clearing goalThreshold. Reports mark this completionBasis as model unless a verify callback also passes, in which case it is model+assertion.
- ModelRouterMiddleware
- An experimental middleware hook where Jev analyzes the latest user input to select which chat model processes the turn.
- n8n-nodes-jev-classification
- MIT n8n community node from khmuhtadin. Classify, Score, Check, and Ask Questions against TypeSafe's Jev. Fixed classify mode emits one output per category plus Needs Review.
- Native computer use
- Clicking and dragging controls in a real macOS window rather than driving a browser DOM.
- native-binary JevBench
- The 74 public text items in this card that are already yes/no decisions. The card says they are not the 534-task composite, the private tests, or the interactive bench at jevbench.dev.
- nDCG@10
- The README's ranking score. Higher means the relevant documents sit closer to the top of the first ten. NFCorpus moves from 0.322 to 0.377. SciFact moves from 0.687 to 0.745. The footnote says these are complete text-retrieval sets and that desktop timing varies.
- NDCG@10 here
- cheat-at-search keeps ranks through 10. Gain is (2^grade - 1) / rank. The ideal is a perfect top 10 at the dataset's maximum grade, the same ideal for every query.
- Needs Review
- The extra classify output when confidence is below the node's threshold, default 0.5. Items still carry jev.needsReview when the node is told to send them to the best category anyway.
- New sources
- Palmer's out-of-domain split: datasets and policy rule types Kev was not trained on. The September 20 post puts Kev-8B at 79.6% there against Jev 85.7%.
- Next-token logit read
- Score the allowed answer tokens from one prefill, with no decode loop. SimpleJev reuses the shared prompt's KV cache across questions in the same request.
- NIRNAY 450M
- Eulogik's Apache-2.0 decision checkpoint at eulogik/nirnay-450m. The card says it is a Laya 421M base plus about 30M, and that it serves POST /v1/systemone.
- No state retention
- Venice's claim that it keeps no copy of the state you send or the answers it returns, and that it does not train on that traffic.
- Noul
- TypeSafe's term for a boolean question, returning a probability score between 0 and 1 without a separate confidence field.
- number type
- A Span-02 question that returns a numeric value and a confidence, without listing candidate numbers. The card marks number, reasoning, and images as extensions of TypeSafe's System One API.
- O at 0.5
- A Jev probability of at least 0.5 is marked O. 342 items sit on 0.5 (49, 291, and 2 by section). Marking those as X moves the gaps from 38.3, 11.3, and 36.8 points to 36.7, 11.5, and 37.5.
- OCR label overlap
- On three pages, OCR reproduced 7 of 13, 7 of 88, and 3 of 85 DOM labels verbatim. The README's reading is that a classifier over OCR blocks is choosing among garbled strings.
- October 9 In Brief
- Marina Temkin, TechCrunch, October 9, 2026, 2:41 PM PDT. The piece says TypeSafe raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz, with Sequoia and DCVC. We did not open a filing.
- one box per question
- TypeSafe Box docs, reopened October 16, 2026. Every Box answer is a single rectangle. TypeLLM's 05:44 UTC post says Jev 1.14 cannot return an array of vehicle boxes. This desk's type array POST, Date header 08:19:46 GMT, returned HTTP 400, Invalid request.
- one run
- JEV.md says experiment 12 put this evaluator's run-to-run range at 1.12 points, and that a single run reproduced the median of three at Pearson 0.998. The corpus therefore scores each language once. The same note says 12 of 268 port-check rows came back bit-identical, so the evaluator is not deterministic.
- one sentence per call
- The Jev run puts a single Korean translation in the state. The three chat models scored twenty sentences in one prompt. The note treats that as the remaining difference in conditions.
- one shared screenshot
- The cost table in awlevin's README. One decision, same screenshot and goal. Jev is $0.0002 and 0.13 to 0.38 seconds. Claude Opus 5 on the bare image is $0.032 and 5.2 seconds. The page says Opus read the dates from the pixels and this loop annotates them in code.
- OneHot
- Sqwish's inference and fine-tuning platform. New accounts may start on a waitlist. Admission gives $5.00 of credit, once per email inbox, valid for 30 days. A request that names no model asks for kettei-core.
- Open d1
- Liquid AI's October 7, 2026 name for the open checkpoints d1-3B and d1-omni-600M. The hosted model d1:free stays on the Liquid d1 page.
- Open weights
- A checkpoint whose parameters can be downloaded. Fastino's August 5 guide separates that from a release that also includes training data and training code.
- Open-Jev
- Zefan Cai's LoRA adapters plus a trained decision head. The released 2B and 9B checkpoints use Qwen3.5. The 27B v1.1 package uses Qwen3.8-27B and 15,471,617 trained parameters. A different project from SemIf, which previously used the name OpenJev.
- OpenAI Jev
- This page's name for the search. The product OpenAI posted is the Decisions API, powered by GPT-6 Luna. On October 22, 2026 OpenAI Developers posted general availability. Date headers 02:23:12 GMT and 02:23:13 GMT October 23, two POSTs from this desk returned HTTP 200. On October 6 the Developers post called it a public beta. ChatGPT Jev, on the posts we read, points at the same API.
- OpenBles Firewall 1
- A 31.8 MB MIT adapter on fastino/GLiNER2.5-Decide. On the agent test split it caught 119 of 120 attacks and flagged none of 120 benign look-alikes.
- OpenResearch
- alphaXiv's MIT toolkit for training decision models with coding agents. Linked from the October 30 OR-decision-4B thread at github.com/alphaXiv/OpenResearch.
- OpenRouter Armor
- A beta plugin, posted October 15, 2026, that scores each tool call against prose rules with a decision model before the call runs. The default on the 492-call chart is Jev 1.13 at 90.7%, 446 of 492.
- opponent read
- The Jev request in poker-table: a choice over range buckets, a noul for folding to each candidate bet size, and a noul for continue strength. The README says hole cards and chat are not in that request.
- option-fork
- Sthānika's scoring design that scores each option as its own branch. The lab page says option order cannot change the answer under that design. Sieve-31B's card describes an answer-slot readout at a fixed temperature.
- OR-decision-4B
- alphaXiv's October 30, 2026 System One LoRA on Qwen3.5-4B, trained with OpenResearch on 117,186 examples. Hugging Face card prints JevBench 193 of 231, 83.55%. The launch chart prints Choice 79.7 against Jev 79.6, and sarcasm 53.6 against Jev 70.7. Apache-2.0.
- OSWorld-50
- The 50-task slice on the October 2 CUA Speedrun chart. The printed Jev callout is 9.8% success and 125 seconds of wall-clock time per task. Other agents on the plot are not given printed coordinates in this note.
- Outcome Jev
- A Qwen3.5-4B-Base LoRA that reads the question, the actions, the last tool results, and the final answer, and returns P(correct).
- Paid team share
- Vercel's figure for the fraction of paying AI Gateway teams that called a model in a launch window. The Jev post uses nearly 13% at 24 hours.
- Pairwise Choice ranking
- One two-option Choice per pair, 'which is more relevant', then a Bradley-Terry fit on the probabilities. This desk asked all 28 pairs for eight items in one request.
- Parallel constrained decoding
- Read the prompt once, then score the allowed answer tokens instead of generating a string. Nimble's Python code builds the typed result from those scores.
- Parallel questions
- Submitting multiple questions regarding the same state within one API request, distinct from classifying multiple different emails simultaneously.
- Parallel sampling
- Evaluating multiple declared questions over a single input state simultaneously without sequential token generation.
- Pass rate
- Successful attempts out of five per problem. The authors used this instead of time-to-diagnosis.
- Pass-through
- LiteLLM's /typesafe/ prefix. The proxy forwards TypeSafe paths, including POST /v1/systemone, and returns TypeSafe's response unchanged.
- pass@1
- Whether the task passed on the single attempt that was run. 25/30 with the filter, 22/30 without, on the FrontierHarness 30.
- Passage gate
- Four Noul questions about one retrieved passage, with the include, conflict, or drop decision left in code. TypeSafe's cookbook records that pattern on 81 auth passages. aifabrice's fixed-threshold version lowered NFCorpus nDCG@10.
- phase_b
- The Banking77 JSON at eval/banking77_phase_b.json. Accuracy 0.8792, n 3080, Brier 0.208, ECE raw 0.089, ECE fitted 0.045. The file has no Jev field.
- Pilot Router
- AgentOS's layer that picks a model tier for a turn, from c0 for a short reply to c3 for deep work. The Jev strategy is opt-in. The default remains the local MiniLM classifier.
- Planted defect
- A problem inserted into synthetic text so the tester knows which passages should be flagged.
- Play credit
- The hosted demo signs you in through Standard Agents and gives $0.25 of Jev spend. Local development uses your own TypeSafe key and skips login.
- plpython3u
- The untrusted Python language PostgreSQL needs for this extension. The README says the function performs external HTTPS work inside the statement, and that a rollback does not undo a request already sent.
- pointer head
- The small scoring head that replaces the language-model head. The blog says it is just over a million parameters and compares each option's hidden state with the hidden state at an answer position. The earlier slot head is described as worse.
- PolicyLM-1.7B
- Musubi Labs' Apache-2.0 moderation classifier, a 1.7B fine-tune of BidirLM-1.7B-Embedding. A message plus a policy returns a score from 0 to 1 for each category, up to 16 in one pass.
- port 8420
- The address the README gives while the app is open. POST /v1/systemone matches TypeSafe's shape. The examples send model laya and an API key of local.
- port fidelity
- In the laya-mlx README, whether the MLX port selects the same answer as upstream Laya on a fixed set of 63 validation questions. The README separates that check from accuracy against a human label or against Jev.
- Post-only
- An order that adds liquidity to the order book rather than filling immediately against existing orders, capturing the spread when matched.
- pplx-decider-flash-v1.2
- The 4B distill, Qwen3.5-4B. The card prints Decision Index 0.3 overall 60.05 one question per request, 58.88 joint. Docs price input at $0.01 per million tokens. Alias pplx-decider-flash-latest.
- pplx-decider-v1-27b
- Perplexity's decision model, posted October 1, 2026. The card says it is an Apache-2.0 fine-tune of Qwen3.8-27B with a 250k context. The October 1 quickstart priced the hosted route at $0.04 per million input tokens.
- pplx-decider-v1.1-27b
- Perplexity's October 6 update of the same backbone. The October 7 quickstart prices the Decisions API at $0.02 per million input tokens and names this model. Decision Index 0.3 ranks it first at 62.8. The card's own table prints 61.56 against Jev at 57.9.
- pplx-decider-v1.2
- Perplexity's October 16, 2026 27B checkpoint, Qwen3.8-27B, joint requests. The card prints Decision Index 0.3 overall 68.09 against v1.1 at 62.30. Docs price input at $0.02 per million tokens. Alias pplx-decider-latest.
- Precision-gated automation
- Fit a confidence cutoff on validation for a target precision, then apply it to held-out traffic. Poly's 90% precision bar auto-handled 67.5% of items with Jev and 35.7% with ModernBERT.
- Prefill
- One forward pass over the input. A decision model stops there and does not generate the answer token by token.
- Promotional pricing
- The Gateway model card's free input and output for typesafe-ai/jev. On September 26 the card still printed Free and still printed an end date of September 25, 2026.
- prompt_jev
- In mejiasd3v/pg-jev, a PostgreSQL function that sends text and typed questions to Jev and returns jsonb. The README's modes are a noul probability, a choice, a score, or several questions in one request.
- Prompted probability
- A probability vector the generative model writes as JSON, then LocalJev validates and normalizes. The README says this is not mathematically equivalent to reading logits.
- public 231
- JevBench's public tasks, 18 families, which this repository scores locally. The file says that number is not the JevBench Score composite, which also uses sealed items, speed, and cost.
- public-proxy Intelligence
- The README's name for a score on the 231 public JevBench v1.2 items. The judge tier is private, and Calibration, Speed, and Cost are not in this number.
- PurpleMIST public index
- 46.7 on a self-run of Decision Index edition 0.3, public suite, 140,620 requests, not yet the board's Full score. Separate from the 0.668 African accuracy.
- PurpleMIST-Flash-1.0
- Olaverse's Apache-2.0 decision model, about 7.9B parameters, from Qwen3.5-9B-Base. On the card's African Typed Decisions harness, translated accuracy is 0.668 on 15,050 decisions. Text only.
- Quality-score variance
- How much a judge's continuous 0–1 quality score moved when the agent output was unchanged. LangChain reports Jev at 0.0000149 mean per-case variance.
- Question isolation
- Running a separate encoder pass per noul so other questions in the same request cannot change that probability. jeff defaults to isolating nouls.
- Quyet 1.0
- Chinh Nguyen's family of decision models. Large is a Gemma 4 31B checkpoint, Apache-2.0. The homepage's frozen JevBench v1.6.0 cell is official score 71.39 and Capability 81.7.
- Quyet sizes
- The README lists Large 31.3B, Medium 4.66B, Small 328M, Small-EN 153M, and Tiny 183M. The homepage rounds Medium, Small, Small-EN, and Tiny to 4.7B, 330M, 150M, and 180M.
- Quyet v1.6.0 and later boards
- The homepage's rank 1 is frozen v1.6.0. The October 8 open-weights rank line says official composite 3 at the same Capability 81.7. H2O's October 7 composite for Quyet is 71.4.
- Qwen Research License
- The card's license for this derivative of Qwen2.5-Omni-3B. It permits research and evaluation use. Commercial use needs the permissions in that file.
- Qwen-Audio-Jev
- A Jev decision head on Qwen-Audio-3.1-ASR-Flash-Next. arXiv 2610.21614, submitted October 15, 2026. Choice, Noul, and Score over audio. The abstract prints 78.70% on MMAU-mini and about 20x the base model's speed.
- range_wide
- A statement whose p_true moves by 0.20 or more across its three wordings. The README says those rows can be ranked and should not be quoted as a single number. 45% of statements are marked this way.
- rank agreement 0.99
- The chart caption's summary of how Jev ordered 25 probability phrases against the human survey. The median gap printed beside it is 3 points.
- raw label softmax
- llamajev's probabilities are a softmax over the logprobs of the option labels. The README says they are not TypeSafe's calibrated RLCD probabilities.
- ReAnchor
- The experimental DSPy 3.4 optimizer that fits Boolean thresholds, Score cuts, and Choice weights to a program metric. The release notes print no accuracy, and the patterns README does not print the fitted ask threshold.
- Reference-label accuracy
- Share of emails where the model's top category matches the label Mudholkar assigned. The 96.4% figure is this metric on 1,565 items, including 364 synthetic mails.
- Reference-label agreement
- Share of examples where the model's top answer matches the synthetic label. On the 324-example table that share is 90.12% for Bespoke-Nimble-9B and 93.21% for Jev 1.13.0. It is not calibration against human judges.
- Reject filter
- Skip the slower model when Jev's P(yes) sits under a line set from known positives. On Kumar's page gates the line came from half the lowest P(yes) among real positives, checked by rotating production runs.
- rev
- Steven Yu's MIT harness. It reads option logits from a frozen open model and serves POST /v1/systemone. The default local checkpoint is Qwen3.5-2B at 8-bit. An upstream mode reads an sglang server instead.
- RISK_JUDGE
- The switch, jev or qwen, behind triage, ranking, scoring, and severity. The Qwen arm answers the same questions in strict JSON. The article reports zero parse failures.
- RISK_TRIAGE_THRESHOLD
- The Noul cutoff that decides whether a sweep escalates. The article ships 0.5. At 0.7, Gulf scored 0.58 and the run stopped, after the 0.5 arm had marked that profile critical.
- RLCD
- Reinforcement Learning for Calibrated Decisions, TypeSafe's training methodology intended to align model confidence scores with true outcome probabilities.
- Router
- Laya's entry point that detects script and language, then dispatches to english, multilingual, or typed-decisions. Lazy load without preload can cost 7 to 10 s on a language switch.
- RoutingPolicy
- The sample in the README. Confidence under 0.60 escalates. Confidence at or above 0.85 is automatic. The band in between asks first. The page says the numbers mirror TypeSafe's guidance and are a starting point.
- Runnable
- LangChain's standard unit of composition that can be executed, batched, or streamed across agent pipelines.
- S1MB Japanese Task Avg
- System One Mosaic Benchmark Japanese average in the paper: chance-corrected scores averaged within each answer type, then across Noul, Choice, and Score, times 100. DF-JA-Large prints 67.63 after specialization.
- Sage 1.3.0
- Levanto's hosted decision API. Opened October 11, 2026, the JevBench API board ranks it first of 27 at composite 74.0. The base on that row is undisclosed.
- Sage output price
- Levanto's pricing page lists output at $10 per million tokens on every plan, and says those tokens are reasoning. It says answers are free. Pro input is $0.042 per million tokens.
- SageCraft
- An open harness that shows Sage a screenshot and a short list of moves. The October 9 post logs 4,573 decisions and level 5 at 2 cents a kill. That log is not a JevBench score.
- samples=1
- The README's fast latency setting, about 35 to 60 ms. The same page says samples set to auto is about 160 ms. Neither cell is a JevBench score.
- Schema
- The structured Choice definition representing possible piece landings, preventing the model from outputting invalid moves.
- Score
- A numeric evaluation based on a custom rubric, returning an estimated position, distribution, and confidence level.
- screen parser
- The thread's account of the extra step in front of text-only Jev. About 500 ms a step, against an image-model prefill the author puts under 200 ms. The model call itself is put under 0.2 seconds.
- sealed 308
- Private decisions added in v1.4. Only system-level aggregates are published. They contribute 20% of Intelligence after a chance correction. Jev 1.13.0 is 86.6% on the public 534 and 36.7% on these 308, a gap of 49.9 points.
- second reading
- When the first pass is unsure, rev scores the options again in reverse and averages them. Confidence stays at the lower of the two readings, so a 0.9 gate still reflects one reading.
- Seeded canvas
- The fixed text of the answer is written into the diffusion canvas before the forward pass. One denoise step then returns a distribution at each noisy slot.
- Selectivity
- The fraction of rows a filter is expected to keep. Malis asked Jev for this number and let Postgres plan from it.
- seller
- Choice returned by this desk's three October 15, 2026 jev-1.14.0 calls, Date headers 11:04:57 GMT to 11:04:59 GMT, on SosyalDijital's published buyer sentence. Confidence 0.48, 0.43, and 0.52. Matching jev-1.13.0 calls chose seller, buyer, then seller.
- Semantic route
- An HTTP route handler matched by a descriptive phrase (such as 'a request from an AI agent') instead of an exact method and URL path.
- semgate
- mizutani's Go net/http middleware. It asks Jev typed questions about an incoming request and hands the answers to your function, which may block, route, or call next.
- SemIf
- Theodore Lee's open project, formerly OpenJev, that reads typed option probabilities from a local model. JevBench v1.2.2 reports 74.6 for the Qwen3.5-4B row.
- Set A
- Per-headline questions: a desk, a story type, and six 0-4 scores (magnitude, velocity, novelty, window, heat, risk) that feed a newsworthiness mark out of 10.
- Set B
- Per-company questions on standing and journalist shape, then a tier. A story can land on several of the 15 illustrative desks, or none.
- settle
- After each action, dejevu waits until the DOM is quiet and the requests that action started have finished, with a cap per action type. The README contrasts that with fixed waits, and says jev-ultrafast discards 4 to 6 of 17 calls because the page changed under them.
- Shadow mode
- Classify and log the label, then act as if control_plane is off. The PR default for live.classify.shadow is true.
- Sieve
- A filter placed before a larger model to discard unnecessary context according to calibrated criteria.
- Sieve-31B
- Sthānika's LoRA adapter, rank 128, on google/gemma-4-31B-it. Hugging Face sthanika-ai/Sieve-31B, license apache-2.0. Serves choice, noul, and score. The card prints Decision Index 0.3.2 Full score 63.8.
- signup pause
- TypeSafe's September 22 statement that new Jev accounts were temporarily not being created, while accounts that already signed up kept working. On September 27 at 22:30 UTC the same account posted that signups were open again, and that new signups no longer get free credits. On October 6, 2026, at 17:05 UTC, a new console account on this desk showed available credits of $0.00, with no payment method and no billing history. On October 8, 2026 the same signed-in console still showed available credits of $0.00, no payment method, and no billing history. Usage for the last 7 days was $0.00, 0 tokens, and 0 requests.
- silence gap
- On a paired 200-item MMAU diagnostic, ToddlerJev scores 73.50% with full audio and 38.50% when the signal is replaced by silence.
- SimpleJev
- Featherless's Apache 2.0 classifier server. It reads next-token logits for allowed labels and returns Choice, Score, or Noul JSON. JevBench v1.2.8 lists Qwen3.8-27B at 67.3.
- single, binary, ordinal
- Sqwish decision kinds on the quickstart. single picks one named outcome. binary is yes or no. ordinal returns probabilities over ordered levels plus expected_level counted from 0.
- sink, don't sort
- postgres-search keeps a top-10 hit in place when its Noul is at least 0.30, and moves a lower score to the bottom of those 10 without reordering either group. The docs say a correct product scores near 1 for close variants too, so sorting by the probability would shuffle good rows.
- Sistema 1
- Coronado's name for the typed path. Six closed questions in one POST /v1/systemone, probabilities only, no generated text. On the realtime table, N=5, mean progress is 24.9% at 408 ms a decision.
- six
- Choice on this desk's October 16, 2026 jev-1.14.0 call, Date header 08:19:46 GMT, on a 960 by 540 frame from Elon Musk's flying-cars clip. six at 0.59, six 0.69, seven_or_more 0.27. TypeLLM's cookbook at 14.00 s prints six vehicle boxes. Those two sixes are different hosts.
- Skill file
- A SKILL.md instruction file read by coding agents. TypeSafe's file prompts agents to fetch updated documentation rather than relying on stale parameters.
- SOAM
- State once, ask many. basal-1.5 reads the state once and scores every question in the request from that pass. On the H100 comparison, five questions take 37.8 ms instead of 111.7 ms.
- social_nicety
- Choice returned by this desk's six October 15, 2026 calls, Date headers 14:57:06 GMT to 14:57:09 GMT, on うちの子ぉも、ええ子にしてますえ. Three jev-1.14.0 calls at confidence 0.69, 0.66, 0.64. Three jev-1.13.0 calls at 0.33, 0.36, 0.34. Yamashita's expected label on that phrase is warning.
- Span
- A Vela 2.0 answer that returns a label, character offsets, the text, and a probability. The 9B card says offsets are Unicode code points, start inclusive and end exclusive. The 0.3B card ships the router span head. The three larger cards also ship a broad span head.
- Span-02
- Respan's decision checkpoint, posted October 15, 2026. Fine-tune of Gemma 4 26B-A4B. The card prints Decision Index 0.3 text index 63.62. The blog prints thinking off at 61.2 against Jev 1.13.0 at 58.0.
- spark-s1-4b-v6
- Abhishek Rai's Apache-2.0 LoRA on Qwen3.5-4B. It restricts the next-token logits to option labels. The README's public-proxy Intelligence is 83.1, which is not the official JevBench composite.
- Spearman 0.74
- Rank agreement on 578 held-out terms that have autocomplete clues, model dated 19 September 2026. Mean error on that row is 10.4 points. A random 20% split on the same page is 0.45.
- spot bank
- 46 handwritten Hold'em spots in the poker-table repo. The Jev mode repeats them three times through OpenRouter on typesafe/jev-1.13. The README gives a median near one second and does not print an accuracy percentage.
- SR-Box
- The d3-ultra card's private Box suite: 5,678 questions on 4,984 images. Ten-domain mean for d3-ultra is 58.2%. d3-turbo on that table is 60.0%.
- SR-Text
- The d3-ultra card's private text suite: 10,411 requests and 73,495 questions. The printed mean of six chance-corrected parts is 87.2 for d3-ultra.
- SREGym-Lite
- A smaller SREGym problem set meant to run on modest clusters. This experiment used 10 of those problems, not the full 21-problem lite suite.
- Stagehand
- Browserbase's library for automating remote browser sessions through structured interaction steps.
- Starting credit
- The $5 TypeSafe posted for every new account on September 20, about 120 million tokens at $0.042 per million input. On September 27 the company said new signups no longer get free credits. On October 6, 2026, at 17:05 UTC, a new console account showed available credits of $0.00. On October 8, 2026 the same signed-in console still showed available credits of $0.00.
- StartLux-Decision
- StartLux Labs' typed decision family. Five dense sizes from 0.8B to 27B, plus a 35B mixture with about 3B active. The server speaks POST /v1/systemone.
- State
- The input data provided to Jev, such as a single email or batch. The calling application retrieves and supplies the data.
- State once
- The card's name for reading the input once per request and answering every question from that read. On the printed MI325X table, 255 questions about one photo take 2.12 s for d3-turbo against 32.94 s for d3-flash.
- state shuffle
- Jevre keeps each row's candidates and label and replaces the state. Test accuracy falls from 0.9323 to 0.4282. Held-out sources fall from 0.8366 to 0.4369. The README presents that drop as a control.
- status.typesafe.ai
- TypeSafe's public status page for api.typesafe.ai and console.typesafe.ai. The September 21 incident is titled Console is unavailable.
- Step Jev
- A Qwen3.5-4B-Base LoRA that scores one observed search step. The score used in training is P(positive) minus P(negative).
- StreamDecisionBench
- Jacob Lin's October 2, 2026 chart of the decision in force while the state keeps changing. The single-model score is normalized log-AUC from 0.5 to 8 seconds. The site HTML we opened did not render the table. The code link in the thread points at the same site.
- strict_mode
- JevEval setting that returns 1 only when every applicable question is answered in its best way from Jev's probabilities, and 0 otherwise. Invented-breeze 0. Clean 1.
- Stripe shortlist
- Eight passages this desk wrote about Stripe keys or webhooks, including Roll secret, the signing-secret page, HMAC verification, and a trap that rotates the API secret key instead.
- Stripe ticket
- The October 6 desk sentence Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP. On October 13, 2026, at about 03:00 UTC, Snap returned urgency noul 0.98, 119 input tokens, 0 output, desk clock 1,195 ms, then chose technical at confidence 0.90, score 0.98, noul 0.98, 235 input tokens, desk clock 2,291 ms.
- Structural invariant
- A relationship you might expect between two questions, such as P(yes) on a Choice matching a Noul, or a Noul and its negation summing to 1. The docs say jev-1.13 does not guarantee those identities.
- Structured read
- OpenJev's path, as LocalJev describes it: a one-step DiffusionGemma pass that needs seeded canvases, read-only denoising, and selected-token logprobs, which stock oMLX does not expose.
- student graduation
- In Mastracci's October 27 thread, the local student trains on EmbeddingGemma2 embedding vectors until it is confident it matches the teacher, then checks in only occasionally.
- SubQ Snap
- Subquadratic's hosted decision model. Docs opened October 23 name subq-snap-1.1 at $0.04 per million input tokens, output free. GET /v1/models listed subq-snap-1.1 and subq-snap-preview. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs from this desk returned HTTP 200. The October 12 blog prints Decision Index 0.2.1 at 61.7 against a Jev cell of 57.9.
- subq-snap-1.1
- Pinned Decision model id on the Models page opened October 23. Whedon wrote that the preview slug routes to it. A POST with that id on October 23 named subq-snap-1.1 in the response. Artificial Analysis Decision Index prints 65, tied with d1-ultra and Elephant 9B.
- swift-jev
- Daiki Matsudate's MIT Swift package and CLI. Choice answers are enum cases. Noul answers are probabilities. Score answers can fall between rubric levels. One evaluate call sends every question.
- System One
- TypeSafe's term for a model returning structured decisions with probabilities rather than text output, inspired by Daniel Kahneman's distinction between fast and slow cognition.
- Target element
- The specific page element receiving the selected action. The repository states that both heads share a request to avoid mismatched actions and targets.
- Task 1
- Given-aspect valence and arousal on a 1 to 9 scale. Ten corpora, six languages. The headline is micro RMSE over every annotation. Lower is better. The paper's number is 1.0645.
- teacher average
- The blog's name for the reference on typed-decisions: the mean of three samples from teacher models. It says two single samples agree 73.5% of the time, so a model can match the average more often than one teacher matches another.
- ten-call clock
- October 7, 2026 desk clocks on the Stripe sentence, ten measured calls per host after one warmup. Urgency alone: Jev median 330 ms, Decisions median 267 ms, openai-processing-ms median 64 ms. Three quickstart questions: Jev median 338 ms, Decisions median 269 ms, processing header median 68 ms.
- Tev1
- Together AI's experimental Qwen3.5 decision model, in 4B and 0.8B sizes. The Together chat call returns one letter. Ollama's library page serves it as Choice, Score, or Noul.
- three-set mean
- The fourth column. It is the average of the three printed percentages. Jev 80.8%, Bonsai 2 67.2%, Luna 85.4%.
- Threshold
- The minimum score required for a match (defaulting to 0.5). The first description whose Noul probability exceeds this value triggers its handler.
- THX-01
- HAL-X's 322M decision model, Apache 2.0, on mmBERT-base. The card's support-ticket average is 98.4 against Jev 1.13 at 97.4 on 2,843 tickets.
- THX-01 API
- The docs page names POST https://api.hal-x.ai/v1/systemone and model thx-01. The October 11 post calls the API free. The docs page we opened does not print a price.
- THX-01 number and excerpt
- The card's question types add number, excerpt, and cite beside choice, noul, and score. Number lookup on the card is 93.4. Jev 1.13 on that line is 96.6.
- tier 0
- Hard denies in the offline eval that do not need a model. 12 of 22 catastrophic commands in the 80-entry corpus were caught there. The other 10 were caught by the degraded ask path.
- ToddlerJev
- Teenage Works' open-weight audio decision model, Hugging Face TeenageWorks/Toddler-Jev, Qwen Research License. Posted October 18, 2026. MMAU-mini 68.70%. Base Qwen2.5-Omni-3B with a rank-16 LoRA and a pointer head.
- TokenBreak
- HiddenLayer's 2025 attack that edits how words split into tokens. The October 14 Jev write-up used a denser, targeted form on 30 toxic prompts. At baseline both Jev and the analog flagged all 30.
- top 128
- After the local word match, JEV reads at most 128 candidates and moves the ones that fit the query. Files it does not check stay in the list as local matches. A search with no API key never makes that call.
- tree search
- PageIndex builds a hierarchy of titles and summaries. Jev picks a child at each level. The script keeps a beam of 3, then runs one Noul per candidate page, up to 16.
- TrueForge
- TrueFoundry's MIT-licensed agent harness. npm @truefoundry/trueforge 0.178.0, opened October 14, 2026, requires Node 22.14 or newer. Local mode is one process on SQLite.
- Two-week paid-team reach
- The October 2026 Production Index launch table's 35.3% of paid teams in Jev's first 14 days, gateway launch Sep 17. TechCrunch rounds that cell to 35%. 0.27% of spend in the summary, 0.3% on the scorecard, and 0.15% of September spend are three prints.
- TypeSafe-Danger-Zone-
- Prefix LiteLLM forwards to TypeSafe deployments only, opened October 15, 2026. TypeSafe-Danger-Zone-Images: true turns on images and box. TypeSafe-Danger-Zone-Long-Context: true raises both caps to 262k.
- TypeSafe-Danger-Zone-Images
- HTTP header, value true, required on every TypeSafe request that contains an image or a Box question. On October 15, 2026, Date header 01:22:14 GMT, a request with an image and no header returned HTTP 400.
- TypeSafe-Danger-Zone-Long-Context
- HTTP header, value true. Opened October 15, 2026, the models page sets both the per-request context and the per-question context to 262k tokens when this header is present. It works on Jev 1.14 and Jev 1.13.
- typesafe/jev-1.13
- The OpenRouter model string main.py passes when use_jev is true. The script does not print a served version such as jev-1.13.0. The checked-in default is use_jev = False. results.json contains one true run and one false run.
- typesafe/jev-1.14
- OpenRouter model id for Jev 1.14. Opened October 15, 2026, the page lists $0.042 per million input tokens, output free, and a 262K context. GET /api/v1/models did not list this id. This desk did not call OpenRouter.
- typesafe/jev-router
- OpenRouter's chat model id for Jev Router. The guide says Jev scores the conversation and the router picks a model and a reasoning effort. The response is text from the model that served the request.
- TypeSafeClassifier
- A LangChain Runnable that submits state alongside defined Choice, Score, or Noul queries to Jev, returning structured results rather than text generation.
- v1.5.4
- The JevBench page we opened September 30. It numbers 106 systems. Cygnet stays 73.7, Winnow-12B Q8 stays 73.2, and Jev 1.13.0 stays 72.1. Former addenda now have official ranks and keep the addendum tag. The capability control says 60 Jev-class systems.
- v19
- The reference Strands Decider recipe in the repository. At the 3072-token window it was preregistered at, the file prints 167 of 231. Served at 4096 tokens the same note prints 168. v20 scored 169 at 4096 and did not replace it.
- Vega, Decision 2.0
- A different checkpoint, Decision-2.0-Vega-27B from vLLM Semantic Router. Its card prints Decision Index 0.2.1 at 56.5. The write-up is the Decision 2.0 page.
- vegaml
- NandhaKishorM/vegaml, Apache 2.0. Default checkpoint 0.8B on frozen Qwen3.5-0.8B, plus a 4B. The README's phishing row is 75.4% against Jev 1.13.0 at 61.9% on 800 emails.
- Vela 2.0
- Four Apache-2.0 checkpoints from KR Labs and vLLM Semantic Router, posted October 7, 2026. A call can return Choice, Noul, Score, Span, or Set. The model does not write a reply.
- Verbatim keep
- Retaining original tool outputs unchanged in context without summarization, preventing drift in variable names or error logs.
- verdict
- act, mark, or fall_back, chosen from thresholds in the jevel. act means the probability clears the bar to use the answer. mark means use it and flag it. fall_back means do what the code did before the call.
- verified run
- In dejevu, a run counts only when code checks the final page. The README says the model announcing that it is done is not the check. Flights must show results. Wikipedia must land on the exact article URL.
- visual inputs
- Tibo Sottiaux's phrase on September 29. The OpenAI Developers follow-up says to send text or images as context. The October 6 guide says images must be inline base64 data URLs. The GPT-6 Luna model page lists image as an input modality and marks image generation as not supported. This desk did not send an image.
- VoxCeleb1-O EER
- Equal error rate on a speaker-verification split. Qwen-Audio-Jev prints 6.50% for its Noul setup.
- Waitlist
- TypeSafe's early-access queue from launch through September 20, 2026, when the company posted that Jev was open with no waitlist.
- WaterSheep
- Samrat Dutta's Apache-2.0 fine-tune of ModernBERT-base. The README serves POST /v1/systemone on port 8766 and adds a multi-label question type.
- web context
- In the Exa demo, search snippets placed into the state Jev judges. The page says Jev is asked the same yes/no question with that state and without it. No scored pair was printed on the page we loaded.
- weekly-client-follow-up
- Akshay Pachaar's TrueForge agent that shortlists 10 GitHub issues per client request and asks Jev which issue matches, or none. Accept at 80% confidence, 90% if the issue is closed.
- Werdykt
- basal's hidden benchmark, v1, exported October 4, 2026. 5,000 decisions, 10 categories in Polish and in English, 250 questions a cell. The questions are not published.
- Werdykt v1
- basal's hidden 5,000-decision set, 10 categories in Polish and in English. The ultra card prints 0.879. The questions are not published.
- Whole-document read
- A question whose answer needs the page as a unit, not a short snippet. Kumar's 135-row extraction gate is the example: Jev matched later outcomes 68.6% and the probability swung by 0.5 between runs on the same page.
- Workers AI
- Cloudflare's serverless inference runtime, executing queries via env.AI.run('typesafe/jev', { state, questions }).
- WrangleBench
- John Crown's chart of perfect matches per 200. The printed bars are 137.3 with Jev and 105.1 without. The post says the full benchmark is still to come.
- x_celeris.explain
- Optional Celeris field that returns one sentence per question. Docs say explanations make the request slower. Output tokens, including those sentences, are free.
- x402 wallet
- A prepaid USDC path Venice lists next to a Bearer API key for authenticating decision calls.
- Zero-probability miss
- The model assigns probability 0 to the correct label. Poly reports this on 6.6% of Jev's decisions and on none of 1,155 decisions from each of the three open encoders. A zero cannot be recovered with a threshold.
- ZeroNSFWBot
- SeYeD.Dev's Telegram bot. Pictures are scored on the host. Jev reads clipped profile and message text when a key is set.