Updated
hunch runs EmbeddingGemma2 as a CPU System One endpoint at 20 ms on text and 100 ms on images
Matt Mastracci's hunch turns EmbeddingGemma2 into a CPU System One endpoint at 20 ms on text and 100 ms on images. Cloudflare Clef teaches a local student in real time. Open-source launch was planned for later the same week. This desk has no public repo or hosted call yet.
On 492 tool calls, OpenRouter Armor's default Jev 1.13 scored 90.7%, 446 of 492
On 492 tool calls, OpenRouter Armor's default Jev 1.13 scored 90.7%, 446 of 492, at $0.0021 a call. On October 16, 2026, at 01:18 UTC, this desk's jev-1.13.0 noul was 0.98 on git push origin main and 0.01 on ls ./reports.
On DeepEval's Paris weather example, jev-1.13.0 scored 0.254 when the answer invented a breeze
On October 14, 2026, at 01:58 UTC, two POSTs to jev-1.13.0 scored DeepEval's invented-breeze case at 0.254. The matching clean answer scored 0.901. Sumanth's October 13 post points at JevEval, DeepEval's Apache-2.0 metric that maps Jev probabilities through a fixed weighted mean. The docs walkthrough that prints 0.461 is illustrative. This desk did not install deepeval.
On 40 client requests, Jev and Opus both scored 37 of 40, at $0.0015 against $0.005 to $0.015
Akshay Pachaar's TrueForge agent scored 37 of 40 client-issue matches with Jev and 37 of 40 without it. The printed decision costs are $0.0015 and $0.005 to $0.015. Review caught 5 wrong Jev matches. On October 13 he wrote that 40 is too small to establish how reliably confidence routes errors to review. On October 14, 2026, at 02:34 UTC, this desk's jev-1.13.0 call picked the intended issue at confidence 1.0. A trap shortlist scored 0.51. This desk did not install TrueForge.
Opened October 26, Decision Index 0.4.1 Full score for d3 is 67.15; d3-turbo and d3-ultra have their own pages
Opened October 26, 2026, Decision Index 0.4.1 Full score for d3 is 67.15. The free API aliases vllm-sr/d3 to d3-turbo. Sibling pages cover d3-turbo at 67.31 and d3-ultra at 70.18. The October 16 card reading of 66.3 on edition 0.3.2 stays.
Vela 2.0's 9B card prints 41.63 on Decision Index 0.2.1
The Vela 2.0 9B card prints Decision Index 0.2.1 at 41.63 with Noul calibration on, and 41.09 with it off. The same evaluation file prints the Lux base at 46.23 and does not print a Jev row. Four Apache-2.0 checkpoints also return span offsets. We did not run the weights.
fsdecide scores 91.3% on ForkScape review, and Jev scores 56.3%
ForkScape posted fsdecide on October 5, 2026, a 149M ModernBERT trained on the app's own decisions. On 80 review items it scores 91.3% against jev-latest at 56.3%. On the authors' 1,248-request general set, jev-latest is 90.5% and fsdecide is 60.5%. That set is not Florian's JevBench.
basal-1.5 scores 72.12% on Werdykt, and Jev 1.13.0 scores 81.64%
Remek Kinas posted basal-1.5 on October 4, 2026. Werdykt v1, exported that day, puts the 4.5B at 72.12% and Jev 1.13.0 at 81.64%. A closed set of 3,000 new Polish decisions is 93.1% against basal-1.0 at 87.4%. The site schedules the weights for October 5. Opened October 8, 2026, GitHub main titles the README basal 1.5 and says the models are on Hugging Face from 2026-10-05.
JEV-27B-VL finishes 15 of 20 pick-and-place scenes
AutoTrust posted a JEV-27B-VL follow-up on October 4, 2026. The model card's October 3 section prints 15 of 20 pick-and-place scenes, and 95% of 60 browser tasks when the element text is included. A six-benchmark mean from September 27 is 84.07 against a hosted Jev cell of 83.85. That JevBench column is not the official board.
StartLux-Decision v1.1 27B Instant prints Full score 69.12, first of 174 on Decision Index 0.3.2
The StartLux-Decision README, opened October 20, 2026, prints v1.1 27B Instant at Full score 69.12, first of 174 on Decision Index 0.3.2. The October 3 self-run of edition 0.2.1 stays 63.88. Code is Apache-2.0. Weights are CC BY-NC 4.0.
Decision 2.0's Vega card prints 56.5 on its own Decision Index run
vLLM Semantic Router posted Decision 2.0 on October 3, 2026. Six Apache-2.0 checkpoints run from Kai at 0.6B to Vega at 27B. Vega's card prints 56.5 on a reproduction of Decision Index 0.2.1 and does not print a Jev row. Its JevArena cell is 74.0, which the card calls level with AutoJev-27B at 72.1. Vela 2.0, posted October 7, is a later fine-tune and has its own page. Nandha Kishor M's October 10 package named Vega is a different model. Decision 3.0, posted October 11, has its own page. d3's card prints 64.7 on its Decision Index 0.3.1 table.
A CUA Speedrun chart prints Jev at 9.8% on OSWorld-50, 125 seconds a task
Pranjal Aggarwal posted a CUA Speedrun chart on October 2, 2026. The printed callout is Jev at 9.8% success on 50 OSWorld tasks, 125 seconds a task. The chart's own annotation says that is as fast as Astra xhigh at 9x lower accuracy. The opening post says 1.5x slower than Opus-5.5-xHigh and 10x lower accuracy. Those two comparisons are not the same sentence.
StreamDecisionBench puts Kev-27B at 62.15 and Jev at 59.63
Jacob Lin posted StreamDecisionBench on October 2, 2026. The chart's single-model leaderboard is normalized log-AUC over 0.5 to 8 seconds. Kev-27B is 62.15. The Jev API is 59.63. The live page's HTML did not render those numbers. The October 1 paper table is a different scoring.
TechCrunch prices one Jev monitor at $2.94, and the demo page prints another estimate
Tim Fernholz, TechCrunch, September 30, describes Shapor Naghibzadeh's Jev Sentinel demo and writes that this kind of monitoring costs $2.94 with Jev against $372 with a frontier LLM, in theory. The demo page does not print those two figures. It prints about $55 per million actions with Jev against about $6,900 for a frontier judge, a 0.23 second median, and 0 refusals on 53,870 attack payloads. We did not open the GitHub repo.
RLCDAlignBench puts a generic Noul at median AUROC 0.886, and median ECE at 0.168
Ruoqi Guo, Yi Liu, and seven coauthors posted RLCDAlignBench on September 24. The set is 10 failure types, 44 benchmarks, and 7,193 instances. Every run in the paper we read uses jev-1.13.0. A generic Noul has median AUROC 0.886 on 31 benchmarks. Median ECE is 0.168 against a null of 0.074, so a threshold fit on one benchmark does not travel. Code is MIT. The dataset is CC BY-NC 4.0 and the Hugging Face copy is gated. We did not rerun it.
Decision Index 0.2 prints Jev at 51.67, and a later post names another first place
Opened October 8, 2026, the Decision Index space still headers v0.3, 115 models, updated 2026-10-07, and Perplexity Decider v1.1 is still first at 62.8. Opened October 7, 2026, the Decision Index space was edition 0.3, 112 models, and its full-score table ranked Perplexity Decider v1.1 first at 62.8. Jev 1.13.0 is a reference at 60.1. Rank 2 is Fastino GLiDE no-thinking at 60.2, and rank 3 is Torchcast Decision 27B at 59.9. apolinario/decision-index reproduces Decision Index 0.2, dated September 24: 40 benchmarks, chance-corrected, five equal areas. A test in the kit locks Jev at 51.67, AutoJev-27B at 50.94, Hopper at 30.01, and Verdict at 1.82. Posts the next day say Drex leads at 51.73. That 51.73 is not in the test we read. The board file generated September 28 is edition 0.2.1 and lists Jev 1.13.0 at 57.91 among 71 names, with no Drex. Nace's page prints Drex 1.5 at 58.28. Maincode's September 30 card prints a self-run of the 0.2.1 kit at 59.26 for Matilda, against this board's Jev cell of 57.91. The card says that run is being submitted, so 59.26 is not a row in the September 28 file. Nace's September 30 post also calls Drex 1.5 first on JevBench. The chart on the next post prints 201 of 231 for Jev and 199 of 231 for Drex 1.5. Liquid's September 29 chart is a separate reproduction at 58.9 against 57.9. Fastino's October 1 blog prints GLiDE at 64.81 on a self-run of the 0.2.1 scorer, against this board's Jev cell of 57.91. AutoTrust's October 3 card prints GEV-26B-Decide at 62.48 on another self-run. Neither figure is in the September 28 file. StartLux's October 3 README prints 63.88 for StartLux-Decision-27B on a self-run of the same scorer, higher than that Jev cell on 31 of 38 benchmarks, and says the run is not on the board. vLLM Semantic Router's Decision-2.0-Vega-27B card, opened October 5, prints 56.5 on its own 0.2.1 reproduction. The card does not print a Jev row, and 56.5 is not in the September 28 file. Pixilab's Blink v0.3 card, opened October 12, prints 57.48 on a self-run of the 0.2.1 kit, against the board's Jev cell of 57.91, and says the run is not a board submission.
Eikos-27B prints 82.9 on its JevBench hard row, and 2.4% error once confidence is 0.90
Caio Vicentino released Eikos-4B and Eikos-27B under MIT. His harness, same items for every system: public JevBench hard tier 82.9, 72.1, Jev 73.0, Laya 35.1. At confidence 0.90 or higher, error on 7,140 items is 2.4% for both Eikos sizes and 5.9% for Jev. A later chess screenshot has Eikos at 53.9% over 388 games. A paper-trading file opened at 09:20 UTC on September 26, cycle 626 of 864, has Jev equity at $10,051.22 and Eikos at $9,653.19. Opened September 30, the model card's RuleArena row is 0.50, 0.50, and a blank Jev cell. The repo's only JSON file is a list of eleven label ids.
CLM-8B caches the action list, and a fine-tuned head scores 31 of 38 DeepSWE holdouts
Jacky Kwok posted CLM-8B on September 23. A frozen Qwen3-8B plus two 20M projection heads. The README says zero-shot latency is up to 9 times lower than Jev. As a verifier, 31 of 38 held-out DeepSWE tasks are correct, and Terminal-Bench 2.1 is printed as 87.6% of 30. Jev, in that setup, sits below pass@1. Apache 2.0.
jevelry turns a Jev call into a folder, a verdict, and a log
Pavel Hegler posted backant-io/jevelry, an MIT npm runtime. A jevel is a folder with one JEVEL.md. The CLI asks Jev once, maps each answer to act, mark, or fall_back, and appends a log. The live sample in the README is one state on jev-1.13.0. Three on-demand API calls took 2.49 seconds.
pg-jev puts prompt_jev() in PostgreSQL and returns jsonb
Jose Mejias posted mejiasd3v/pg-jev, an MIT extension. prompt_jev sends text to Jev from SQL and returns a probability, a choice, or a score as jsonb. The README we read has no accuracy or latency table. Calls are not batched across rows. Zachi's realZachi/pg-jev is a different extension.
An n8n community node routes Jev answers onto category branches, including Needs Review
khmuhtadin posted n8n-nodes-jev-classification, an MIT community node that wraps TypeSafe's Jev as classify, score, check, and ask. One output per category, plus a Needs Review branch under a confidence threshold. n8n declined to verify it because it is building a built-in classifier. Self-hosted only.
MotherDuck's prompt_jev() labels 100,000 AG News rows in 40 seconds for $0.50
MotherDuck shipped prompt_jev(), a SQL function that calls TypeSafe's Jev. On 100,000 articles drawn from the AG News training split, the published table is 2,484 rows per second, 89% accuracy, $0.50, and 40 seconds. GPT-5.6 Terra on the same sample is 52 rows per second, 88%, $37.58, and 31 minutes 59 seconds.
Open-Jev 27B v1.1 answers 197 of 231 public JevBench tasks
Zefan Cai released Open-Jev 27B v1.1, a LoRA and decision head on Qwen3.8-27B. On the 231 public JevBench tasks it is correct on 197, three behind Jev 1.13.0 at 200. The hard tier is 80 of 111 against Jev at 81. The September 21 audit of the released 2B and 9B checkpoints, 150 and 179, is unchanged. The card prints no speed comparison. On October 2 one 64-step pilot on the released 2B checkpoint moved a 128-row synthetic test from 84 of 128 to 99 of 128. Those rows are not JevBench.
A 100-case security triage demo put Jev at 65.3% balanced accuracy against Terra 75.1% and Opus 91.2%
RZ posted a recorded comparison of Jev, OpenAI Terra in Codex, and Claude Opus in Claude Code on 100 synthetic TypeScript security cases, five passes each. Jev's balanced accuracy was 65.3% at 0.57 s median. Terra 75.1% at 10.23 s. Opus 91.2% at 13.83 s. Jev repeated the same answer on 95 of 100 cases. Cost about $0.015 versus $15.02 and $13.94 for 500 decisions.
Vercel's form router keeps Jev's destination when confidence is at least 0.95
Vercel published a Next.js template that asks Jev to pick a team for a form submission and sends the same submission to GPT-5.6 Luna Fast when confidence is below 0.95, missing, or the call fails. The fallback does not see Jev's answer. Tests mock both models.
jevctl puts Jev on the command line, with a compaction hook that keeps lines verbatim
jaidev posted jevctl, a friend's npm CLI that wraps TypeSafe's Jev as shell commands. verify, screen, classify, extract, find, rerank, match, route, ask, compact, and batch. Keys can be TypeSafe, OpenRouter, or Cloudflare. compact drops stale tool results the way Tran's sieve does, at a 0.5 keep threshold.
Jevflake puts Jev in Snowflake: noul, choice, and score as SQL functions
Luke Kranz published Jevflake, a dbt package (and a Terraform module) that lets Snowflake call TypeSafe's Jev API from SQL. Answers are stored once, reused, and tested like any other model. Version 0.1, run against one live account. Row content leaves Snowflake.
semgate asks Jev about each HTTP request, then a live range to attack it
mizutani published semgate, a Go net/http middleware that sends each request to Jev as typed questions and lets your function block, route, or pass. A companion range, Injection Range, is up on Cloud Run. Default configuration sends every header, including Authorization and Cookie, unless you denylist them.
LiteLLM adds day-0 Jev 1.14 images, box questions, and long context on both decision routes
LiteLLM's October 15, 2026 post adds day-0 support for Jev 1.14 images, box questions, and long context on /v1/systemone and /v1/decisions. The proxy forwards TypeSafe-Danger-Zone- headers to TypeSafe only. This desk did not run the proxy.
JevBench v1.5 ranks Cygnet 73.7 and Jev 1.13.0 72.1
Opened October 19, 2026, JevBench JSON revision v1.6.29 ranks 184 open systems and 38 hosted ones. basal-omni-1.0-ultra is first of 184 on Capability at 82.5. kettei-fat stays first of 38 hosted at composite 77.8. Jev 1.14.0 is ninth at 70.2. Opened October 17, the API board named revision v1.7.46 and 39 hosted systems, with kettei-fat first at 77.8 and Jev 1.13.0 seventh at 71.5. Opened October 11, that board ranked 27 hosted systems, Sage 1.3.0 first at 74.0. JevBench v1.5.0, posted September 28, scores 1,624 decisions. The v1.5 page ranks Cygnet 73.7, Winnow-12B Q8 73.2, and Jev 1.13.0 72.1.
LangChain ran Jev as a judge: 500 matching pass/fail labels, and a $0.34 bill against Claude's $28
Daniel Shea and Seán Roche compared Jev with GPT-5.6 Luna, GPT-5.6 Terra, and Claude Sonnet 4.6 on five frozen weather-agent traces. Jev matched a human pass/fail label on all 500 repeated decisions. Continuous quality-score variance was 92 to 913 times lower than the LLM judges. Five cases, one reviewer.
Sam Saffron tries Jev as a hidden control plane for GPT Live in term-llm
Discourse cofounder Sam Saffron opened a term-llm pull request that classifies each GPT Live voice turn with TypeSafe's System One API in about 650 milliseconds, instead of a tool-calling fast-model turn. Shadow mode is on by default. Failures pass the utterance through unchanged.
Charly Poly put Jev next to open encoders, not GPT-5.6, and it lost on calibration
Browserbase's Charly Poly ran Banking77 against ModernBERT, BART-MNLI, and GLiClass. Jev won macro-F1 and automated about twice as much traffic at a 90% precision bar. Expected Calibration Error favored the 396M open encoder, and Jev assigned probability 0 to the right label 6.6% of the time.
Hono's author routes HTTP by meaning, then warns you not to
Yusuke Wada published hono-jev-router, an experimental Hono router that asks Jev which plain-language description matches an incoming request. The README treats it as a demo with sharp limits.
OpenRouter, Cloudflare, and Netlify start serving Jev
Opened October 19, 2026, Cloudflare lists typesafe/jev-1.14.0 at $0.042 per million input tokens, output free, and a 32,000-token window. OpenRouter's typesafe/jev-1.14 listing stays 262K. GET /api/v1/models did not list that OpenRouter id. This desk did not call either host.
Venice lists Jev on its API: same key, no copy of the state
Venice put Jev on a beta Decisions API the night of September 18. Calls go to POST /api/v1/decisions with model jev-latest, billed at TypeSafe's $0.042 per million input tokens with free output. Venice says it keeps no copy of the state or the answers, and TypeSafe quoted the launch.
Vercel's October 2026 AI Gateway Production Index prints Jev at 35.3% of paid teams in its first two weeks
Vercel's October 2026 AI Gateway Production Index, published October 15, prints Jev at 35.3% of paid teams in its first two weeks, nearly double Sonnet 5.5 at 18.1%. TechCrunch, October 16, 15:33 UTC, rounds that to 35%. The September 18 post still prints nearly 13% of paid teams in the first 24 hours.
A week of Jev repos: routers, MCP, graphs, games
A widely shared Awesome Jev thread listed fourteen working patterns; community directories passed a hundred projects within days. The repeating pattern is choose, score, and stop.
LangChain ships a Jev classifier, then parks it on the agent loop
Official LangChain docs add langchain-typesafe, introducing a Runnable for classification alongside experimental middleware hooks for agent loops.