Updated
Desk
People
248 people whose public posts or write-ups became stories on this desk. Handles are how they posted. For how a story gets on the list, see method.
Diogo Almeida
@CompleteSkeptic
TypeSafe CEO. Launch thread, Jevons-paradox posts, and the official System One write-up. On October 14, 2026, at 21:52 UTC, he posted Happy JevDay and jev-1.14.0 as jev-preview. On October 15, at 04:12 UTC, he wrote that the antibench warning is not humor and that TypeSafe does not want people to blindly trust those evals. On September 22 he said new signups are suspended so the team can sleep. On September 27 the company account reopened signups and dropped the free credit for new accounts. He also asked whether Exa's web-context demo was real. On September 29 he posted that the clone war had begun, then wrote that more competition is good for developers if the model is good.
Allie Laabs
@allietheicon
TypeSafe. Named with Diogo Almeida on the October 14, 2026 Jev Day town hall teaser at 1:14 PM.
Sasha Sheng
@hackgoofer
TypeSafe cofounder. Posted that Jev 1.13 is still too slow, too expensive, and too dumb, and pointed at the jaggedness docs.
Gregor Zunic
@gregpr07
Browser Use. The 7-second Google Flights clip and jev-ultrafast.
Ashutosh Mathore
@ashutoshftw
200-piece Tetris bake-off: Jev 9,200 points at about 300 ms per move, 0 illegal landings, against Haiku and Gemini.
Kyle Jeong and Joey Kudish
@kylejeong / @jkudish
Stagehand computer use, plus jev-browser.
Tamara Tran
@tamarajtran
Compaction by scoring tool calls instead of summarizing them.
AstroHan
@AstroHanRay
30 FrontierHarness tasks with a Jev context filter: 25/30 versus 22/30. meriyah 49/49 versus 0/49.
Luke Kranz
@lkkrnz
Jevflake, a dbt package that calls Jev from Snowflake SQL. Version 0.1, one live account.
mizutani
@m_mizutani
semgate, a Go HTTP middleware, plus Injection Range on Cloud Run.
Nandha Kishor M
Convai Innovations
Laya, an Apache 2.0 typed-decision engine. JevBench 70.1. Author table trails published Jev on Banking77. On October 10, 2026 the same author posted vegaml. The README scores 75.4% on 800 phishing emails against Jev 1.13.0 at 61.9%. That package is not Decision 2.0's Vega.
Yusuke Wada
yusukebe
Hono semantic router on Cloudflare Workers AI, then a warning not to let Jev write the reply.
Ryan Vogel
@ryanvogel
1,500 personal emails on a livestream. Sample size is in the post. Accuracy is not.
Jarrod Watts
@jarrodwatts
jev-trader on Monad/Kuru. Defaults to a mock model and dry-run without PRIVATE_KEY.
Mike Taylor
Every
777 writing judgments in under 0.7 seconds, then a 12-passage defect check against Claude Fable 5.1. On September 22 he posted a chart of 25 probability phrases: rank agreement 0.99 with 99 people, median gap 3 points.
Jackson Clark
@HacksonClark
SREGym-Lite: Codex at 20/50 without Jev, 24/50 with jev_plan and jev_submit.
Elvis
@elvissun
News Desk Dealer: 384 headlines in 24.9 seconds at $0.19. Opus 5 finished four.
Mahesh Sathiamoorthy
@madiator
Bespoke Labs. Nimble 9B LoRA, 90.12% versus Jev 93.21% on 324 synthetic labels, and 74.8% versus 76.0% macro on 13 public subsets.
Hemant
@heman10x
Verdict, a 151M encoder that runs in a browser tab. On 337 TypeSafe public cases: 48.1% versus Jev 90.8%. JevBench listed the 1.4 engine at 72.5.
Justin Schroeder
@jpschroeder
JevPilot, a Three.js driving sim that asks Jev which sampled path to take. On October 14, 2026, at 21:07 UTC, he posted that TypeSafe announced image support in the Discord town hall, with TypeSafe-Danger-Zone-Images and TypeSafe-Danger-Zone-Long-Context.
Wes Bos
@wesbos
Posted on October 14, 2026, at 23:15 UTC, that one question on an 850 by 850 image billed $0.000096936, about 10,000 images a dollar, and that referencing the image more than once duplicates token usage. At 23:27 UTC he wrote that 1568 is 4x more expensive than 784 by 784. This desk's 850 PNG used 2,238 input tokens.
Michael Malis
@mmalisper
Postgres JOB overlay. Join-order alone was slower. Hybrid override: 12% geomean.
Nikhil Mudholkar
@nikhilmudholkar
Bryo. 1,565 supplier emails, Jev 96.4% versus two Gemini models a point or two higher.
Charly Poly
@whereischarly
Banking77 against open encoders. Jev won macro-F1 and lost Expected Calibration Error.
Aman Kumar
@onlyoneaman
About 16,000 calls, four public sets, then production page gates. Filter, not classifier.
Logan Markewich
@LoganMarkewich
jeff, a local /v1/systemone server on GLiFormer, with a 1,600-item table against live Jev.
GitHub Next
@GitHubNext
LocalJev, a prompted Bun stand-in for the System One API, plus a 1,200-request bake-off.
Josh C. Simmons
@drjoshcsimmons
Jev macOS Loop. Six native GUI tasks passed. Finder sorted nine files in 7.39 seconds.
Sam Saffron
@samsaffron
term-llm pull request 1150: Jev as a hidden control plane for GPT Live, shadow mode on.
Jake Gwon
@jake_gwon
jevis, a Flutter integration-test package that ranks registered widget actions.
Florian S
@airesearch12
JevBench. v1.5.0 ranks Cygnet 73.7, Winnow-12B Q8 73.2, and Jev 1.13.0 72.1, and calls the top pair a statistical tie. Jev's capability on that page is 80.0. The older v1.4.2.2 harmonic mean still has Imajev-4B at 67.37 and Jev at 63.29. v1.2.8 geometric mean was Jev 75.4. classifier.dev's fast tier is Jev behind another API and is listed, not ranked, on v1.5 at 74.7. The page opened September 30 is v1.5.4 and numbers 106 systems. Those three scores stay on top. The page opened October 4 is v1.5.6 and numbers 110 ranked systems. imajev-4b is official rank 10 at 70.4. The API board opened October 7, 2026, release v1.6.1 under revision v1.7.12, ranks Jev 1.13.0 third at composite 71.5, with capability 77.1. Opened October 8, 2026, revision v1.7.21, that board ranks 26 hosted systems and Jev 1.13.0 is fourth at the same 71.5.
Daniel Shea and Seán Roche
LangChain
Jev as a judge on five frozen weather-agent traces. 500 matching pass/fail labels, $0.34 against Claude's $28.17.
Jeffrey Ip
@jeffr_yyy
Confident AI, DeepEval. On September 22, 2026 he introduced Jev in deepeval. On September 28 he introduced JevEval. PyPI deepeval 4.2.8, opened October 14, lists him as author. On October 13 he replied to Sumanth that JevEval is a step up to using Jev out of the box.
Akshay Pachaar
@akshay_pachaar
October 13, 2026 recap, and the October 12 article, on Jev inside TrueForge. 40 client-issue matches, 37 of 40 with Jev and 37 of 40 without it. Printed decision costs $0.0015 and $0.005 to $0.015. Review caught 5 wrong Jev matches. At 09:21 UTC he wrote that 40 is too small to establish how reliably confidence routes errors to review.
Sumanth
@Sumanth_077
October 13, 2026, 09:43 UTC post on JevEval. The caller writes bounded decisions, Jev returns probabilities, DeepEval calculates the score. Linked the DeepEval blog.
Jason Lemkin
@jasonlk
600 SaaStr Connect judgments. 90× cheaper than Sonnet, 32% false admissions on 148 negatives.
Theodore Lee
@theoleecj
SemIf, formerly OpenJev. Qwen3.5-4B logit readout. JevBench 74.7 versus Jev 75.4.
Jared Palmer
@jaredpalmer
Kev, an Apache 2.0 family of Jev-like models on Qwen. Author new-source table: Kev-8B 79.6% versus Jev 85.7%. JevBench listed kev 0.6B at 66.7. On September 25 he posted Kev-4B on OpenRouter. Jacob Lin's October 2 chart prints Kev-27B at 62.15.
Oleh Aleinyk
@oaleynik
Paired Kev-4B and jev-1.13.0 on September 26. Full JSON function selection is 1,216/1,253 against 1,235/1,253. The historical CLM column on those requests is 47.49% for full JSON.
LiteLLM
@LiteLLM
TypeSafe pass-through on the proxy, plus a Jev compaction guardrail that drops older tool results below 0.2. Kerry Lu and Mateo Wang's October 15, 2026 post adds day-0 Jev 1.14 images, box questions, and long context on /v1/systemone and /v1/decisions. This desk did not run the proxy.
Kerry Lu
LiteLLM
Co-wrote LiteLLM's October 15, 2026 day-0 post for Jev 1.14 images, box, and long context, and the September 20 pass-through note. This desk did not run the proxy.
Mateo Wang
LiteLLM
Co-wrote LiteLLM's October 15, 2026 day-0 post for Jev 1.14 images, box, and long context. This desk did not run the proxy.
Aviz Maeir
@aviz85
Posted on October 15, 2026, at 22:04 UTC, a 16-tweet Box stress test. Same 180 scenes against OpenAI Responses API plus strict JSON: Jev 1.14 82 percent, 0.21 s, $0.15 per 1k, gpt-6-luna 84 percent, gpt-6.1-sol 95 percent. Confidence 0.7 or above was right 107 of 108. This desk did not rerun the 180.
dave
@dave8172
Posted on October 15, 2026, at 23:14 UTC, Where's Pip, a Waldo-style game that asks Jev 1.14 for a Box. This desk did not play it.
稲津 定晃
PC Watch
Wrote PC Watch's October 16, 2026 07:18 JST article on Jev 1.14.0, 22:18 UTC October 15. The piece reprints TypeSafe's blog. No independent table.
alphaXiv
@askalphaxiv
Posted OR-decision-4B on October 30, 2026, at 16:06 UTC. Hugging Face card JevBench 193 of 231, 83.55%. Launch chart Choice 79.7 against Jev 79.6, sarcasm 53.6 against Jev 70.7. Apache-2.0 LoRA on Qwen3.5-4B. This desk did not run the weights.
OpenBMB
@OpenBMB
Posted MiniCPM5-2B-Jev on October 28, 2026, at 13:01 UTC. Project prints 78.79% on JevBench, first among open models at or under 2B, and decision-v7 86.41% with ECE 0.0307. Apache-2.0 on MiniCPM5-2B. Weights ytbai/MiniCPM5-2B-Jev. This desk did not run the weights.
Necati Demir
@ndemir
Posted BranchOne on October 29, 2026, at 13:31 UTC. Open checkpoint returns Choice probabilities from Qwen3.5-0.8B-Base plus LoRA. His A100 Colab median on ten short text questions is 0.305 s. Apache-2.0. Weights ndemirdemir/BranchOne. No published JevBench score. This desk did not run the weights.
Matt Mastracci
@mmastrac
On October 27, 2026 he posted hunch, an EmbeddingGemma2 System One endpoint on CPU at 20 ms on text and 100 ms on images, with Cloudflare Clef as a real-time teacher. Opened October 28, public search found no hunch GitHub repository. Earlier DiffusionGemma-as-Jev clips include phone-camera vision. JevBench tags him on the djev row; the hosted API is Maisa's djev.dev. djev-spark is the server taeold/djev-run deploys to Cloud Run. His September 17 chart of eight hand sets adds to 198 of 201 against Jev at 191 of 201. Pull request 57250, merged September 22, is the vLLM path.
Featherless AI
@FeatherlessAI
SimpleJev, an Apache 2.0 classifier server that reads next-token logits. JevBench listed Qwen3.8-27B at 67.3.
jaidev
@hijaidev
Posted jevctl, a friend's npm CLI wrapping Jev as shell commands. Repo is Nasrallah-AL/jev-cli.
Jerry Liu
@jerryjliu0
LlamaIndex cofounder and CEO. DocJev classifies and splits PDFs with local LiteParse and hosted Jev. Author table: 40/40 classify, 7/8 split, 5.73x / 6.45x versus Luna.
TypeSafe AI
@typesafeai
On October 14, 2026, at 17:16 UTC, posted what's this and a 1:14 PM town hall. The 01:02 UTC October 15 recap says jev-preview points at jev-1.14.0. Posted the September 21 API and console incident, then We are back. On September 22 the same account paused new signups. On September 27 it 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. On October 8, 2026 the same signed-in console still showed available credits of $0.00, no payment method, and no billing history.
khmuhtadin
@khmuhtadinn
n8n-nodes-jev-classification, an MIT community node with one output per category plus Needs Review.
Prasanth J
@prasanth_j
Native C++ DuckDB extension that batches Jev over SQL rows. Live Choice run: 2,049 synthetic rows in 0.887 s, cached replay 28.6 ms.
Kyle McLaren
@kylemclaren
jevsearch, a shadcn site-search block that re-ranks keyword hits with Jev. Author table: 83% Hit@1 versus 41% on the keyword pass.
RZ
@snoopydev99
Recorded 100-case security triage of Jev, Terra, and Opus. Balanced accuracy 65.3% / 75.1% / 91.2%. Repo Robertzu43/system-one-security-triage.
Zefan Cai
@Zefan_Cai
PhD student at UW-Madison. Open-Jev, a LoRA plus a decision head. On 231 public JevBench tasks, 27B v1.1 is 197/231 against Jev 200/231. The released 9B checkpoint remains 179/231. On October 2 one 64-step pilot moved a synthetic 128-row test from 84/128 to 99/128.
Khaled Eltokhy
@eltokh7
jsort, an MIT pairwise ranker. CommonLit r 0.824 at $0.046. Package author in keltokhy/jsort.
Joe Weisenthal
@TheStalwart
Posted a JSort rescore of 4,005 FOMC speeches and statements. The repo's published Fed check is 95 opening statements.
Ben Sabic
Vercel
Author of the form-router guide. The September 21 announcement is the @aisdk post. Jev keeps the destination at confidence 0.95. Below that, Luna Fast decides from the same submission.
MotherDuck
@motherduck
Shipped prompt_jev(), a SQL function on paid plans. On 100,000 AG News training rows the published table is 89% accuracy, 2,484 rows/s, $0.50, 40 seconds.
Chirag
@chirag
Posted llamacpp-jev. The repo is NakliTechie/llamacpp-jev, MIT. It serves /v1/systemone from an unmodified llama-server. Fresh-image median on an idle M4 Pro is 526 ms. NakliTechie/sieve, also MIT, says reversing the options flipped 24% of DiffusionGemma-Jev's answers. The README's September 20 citation is @kis. A reply links a fork that changes llama-server.
きしだ
@kis
The September 19 post that llamacpp-jev cites, 21:43 UTC, shows Qwen3.5-2B answering four questions on a 448 image in about 800 ms. A reply at 22:37 UTC links kishida/llama.cpp branch jev. That branch builds the endpoint into llama-server. The October 3 copy of docs/jev.md lists Qwen3.8 Flash-Next at 0.944 on 1,191 questions. Opened again October 6, the quick start still says the endpoint is not upstream. The newest commit on that file is still October 1.
Xuan-Son Nguyen
ngxson
Opened ggml-org/llama.cpp pull 29818, with Victor Mustar on the October 2 Hugging Face note. The pull merged October 2 and adds POST /v1/systemone for Julia-1, Laya, Kev-4B, lev, and OpenJev. The note's medians are one question on one RTX PRO 6000. Release v0.6.0, published October 5, lists Nimble on that endpoint and Clef with images. Pull 29844 merged October 2. Pull 29831 merged October 3. Pull 29969 merged October 5.
syumai
@__syumai
jevyoumean, MIT, command jym. One Jev Choice against a CLI's help when the subcommand is unknown. Valid commands do not call the API.
Vini Lana
@oviniciuslana
jev-gateway, MIT npm proxy for Codex, Claude Code, OpenCode, and Gemini CLI. Chess-engine bench: Astra and Sol bug-fix output -57%. Opus 5's feature task was slower.
Christian Tzolov
spring.io
Wrote the September 21 Spring blog on spring-ai-community/spring-ai-typesafe. Three-question ticket call, 310 ms median on his laptop. The blog says 0.1.0 is on Maven Central.
andreapn
use-agent-os
GitHub user andreapn opened AgentOS pull request 3316, the opt-in Jev Pilot Router strategy. The announcement chart is 0.901 accuracy against the local classifier at 0.355, on 121 labeled turns.
0xNeoArch
@0xNeoArch
Posted mizorewww/laya-mlx. The README's M3 Max median is 13.42 ms for the 421M checkpoint. Selected answers matched upstream Laya on 63 of 63 validation questions.
maxli
@maxlibin
jev-rubiks, an MIT browser cube. Code solves. Jev decides whether the coach speaks. A removed test left Kociemba distance at 22 after forty move picks.
Ishan Goswami
@TheIshanGoswami
Posted Exa's Jev demo at demos.exa.ai/jev-web-context. The post says web results make yes/no answers more accurate. The page we loaded does not print the pairs.
Latent Node
@latent_node
Posted Decider 1 (sd-1) on meraGPT. The blog has no byline. Its table is accuracy 0.768 against Jev 1.13.0 at 0.727, KL 0.096 against 1.442, on teacher-average labels.
Idov Mamane
@idovmamane
idovmamane/dejevu, MIT. llama-3.3-70b on Groq finished the Flights check in 5.63 seconds. The Jev cell, 7.09 seconds, is jev-ultrafast's published number and was not rerun in that repo.
Jose Mejias
@MejiasDev
mejiasd3v/pg-jev, MIT PostgreSQL extension. prompt_jev returns jsonb. The README prints no accuracy or latency table. A different repository, realZachi/pg-jev, is a separate story.
Pavel Hegler
@pavelhegler
backant-io/jevelry, MIT npm runtime. A JEVEL.md sets act, mark, and fall_back thresholds. The last live test in the README is three calls in 2.49 seconds on jev-1.13.0.
Peter
@LordMarket22
Posted a 612-row banking-feed categorization. Jev 1.13 at 36.1% and $0.023 in 6 seconds. Gemini 3.7 Flash at 42.5%. No repository in the post.
Jinzhengxu
poker-table
GitHub account on Jinzhengxu/poker-table, GPL-3.0. The September 22 post that links it is @bigdicksinapig. Jev reads the opponent. Code picks the bet. Spot-bank median about one second. The post's live pill says 545 ms and $0.002 a hand.
Steven Yu
@stevenyuyy
54yyyu/rev, MIT. A frozen Qwen3.8-27B reads option logits. Public hard tier 0.784 against the README's Jev cell at 0.730. The laptop 2B checkpoint is 0.550 on that tier.
Michael
@michael_chomsky
Posted hotdog-27B, a rank-16 LoRA on Qwen3.8-27B. Independent holdout 1,603/1,631 against Jev at 1,587/1,631. On 74 public binary items the same card has Jev ahead, and the interval includes zero.
Abhishek Rai
@abhi43210
abhishek085/open-spark-jev, Apache-2.0. spark-s1-4b-v6 prints a public-proxy Intelligence of 83.1. On JevBench v1.5 the same checkpoint is official rank 7 at 68.2. On Kev's development partition the README puts it at 0.756 against published Jev at 0.857. The v1.5.4 page keeps 68.2 and lists official rank 13.
Daniel Lee
@dylayed
taeold/djev-run. Cloud Run recipe for DiffusionGemma-Jev on an RTX PRO 6000. The README's samples=1 latency is 35 to 60 ms. Google's Gemma account reposted it on September 22. The repository lists no license.
Mustafa Kendigüzel
@mustafakndgzl
ASO Workbench, a Mac app. Jev answers brand, demand, and fit questions. A local model then estimates Apple popularity. On 578 held-out terms, Spearman 0.74 and mean error 10.4 points.
t.toda
@Trtd6Trtd
Posted a Japanese benchmark table. Jev 96.0% on JCommonsenseQA, 90.2% on JMMLU, 56.0% on JamC-QA. The image prints no item counts.
Daiki Matsudate
@d_date
d-date/swift-jev, MIT. Choice, Noul, and Score in one call. The sample policy escalates under 0.60 and acts at 0.85. The README prints no scored set.
Alex
@alex_forman1
alexforman1/postgres-search, MIT. SQL hit rate 82% first on 50 food queries. The Jev step sinks a top-10 hit under 0.30. The docs say that step was not measured.
Leo
@secondfirst_ai
2nd1st/Jevsus, MIT. The README names Leo as a writer of the sentence bank, with Claude and Codex. September 20 diode on jev-1.13.0: the post counts 209,628 calls and $3.13. 45% of statements move by 0.20 or more across three wordings.
Simone Rizzo
simone-rizzo
Rizzo-AI-Academy/rizzo-flow, Apache-2.0. Spark-X2.5-4B through llama.cpp. authored144 0.812 against SemIf's published 0.819. 6 of 36 missing-evidence rows stay confident. The September 23 post is @QingQ77.
Abel Coronado
@abxda
abxda/ambar-reflex-sistema-1, MIT. A local server plays Ámbar Reflex. Qwen3.5-4B reaches 24.9% of the level at 408 ms. Qwen3.5-9B is 82.7% accurate on 554 decisions and reaches 17.5% on the clock.
あるごす
@argos_M1111
Argos1111/jev_local. Sarashina2.2 Vision 3B on the same API as LFM. JGLUE test: Q4 is 34% on JNLI and 86% on JCommonsenseQA, against LFM at 17% and 69%. The post's usual latency is about 500 ms on a Radeon AI Pro R9700.
awlevin
@awlevin
awlevin/typesafe-computer-use, MIT. OCR, then Jev picks the action. One shared screenshot: $0.0002 against Claude Opus 5 at $0.032. The September 23 post is Somi, @somi_ai. The README says the date parse is code.
Di Zhang
@di_zhang_fdu
trotsky1997/jevre, MIT, 0.85B. Test accuracy 0.9323, ECE 0.0079. Held-out sources 0.8366, ECE 0.0944. Served median 77 ms. The post says RLCD. The README names Plackett-Luce ranking and a Brier term.
vase.park
@thrownvase
Posted a Jev pass over 28,074 translated sentences in hw725/dansa-research. Gaps are +38.3, +11.3, and +36.8 points. The note prints no gold-label accuracy.
Chen
@xinyao27
xinyao27/jevonian, AGPL-3.0. A local proxy on port 8787. jevonian/auto asks Jev once for the model and the thinking depth. The README prints no scored table.
Nodehaven Online
@tbytefrontier
Posted jevgate on September 23. crates.io lists v0.5.1, owner Tauan BF. Findings are composed at probability 0.80. The README says that figure is a judgment, not a measured accuracy.
码农暖爸
@Delroy715
Posted 100 English judgment items. Laya's English checkpoint is 92 correct. Jev is 100. The post says the set is not a full benchmark.
andre
@AssadiAndre
assadiandre/jev-search. macOS file search with no saved index. nDCG@10 moves from 0.322 to 0.377 on NFCorpus and from 0.687 to 0.745 on SciFact. The repo root has no LICENSE file.
七誌
@7shi
7shi/multilingual-reader, CC0. One jev-1.13.0 pass scores 16 models on 67 languages: 1,072 evaluations, $0.3165, 5 minutes 23 seconds. The previous ruler was a Qwen 3.6 median of three runs.
El Asombroso Nosequien
@sancheznoseke
elalber2000/jev-rlm, MIT. One 100,000-line needle. Both arms returned 2867825. The Jev arm took 1,081 seconds. The other took 1,480. The post's dollar figures are an estimate.
Jamie Watters
@Jamie_within
TheWayWithin/jev-bench, MIT, 42 claims. Five runs on the hard 18: Jev 73.3%, GPT-5.4 70.0%, McNemar p = 1.000. He calls the repo JevBench. The 534-task table on this desk is a different project.
Jacky Kwok
@jackyk02
Contrastive-LM/CLM, Apache 2.0. CLM-8B is a frozen Qwen3-8B with two 20M heads. Verifier holdout: 31 of 38 DeepSWE tasks. Terminal-Bench 2.1 is printed as 87.6% of 30. Coauthors on the citation include Christopher Ré and Azalia Mirhoseini.
Caio Vicentino
@0xCVYH
caiovicentino/eikos, MIT. Eikos-27B prints 82.9 on his public JevBench hard row, against Jev at 73.0. Error at confidence 0.90 or higher is 2.4% on 7,140 items, against Jev at 5.9%. A September 24 chess screenshot scores 388 games at 53.9% for Eikos.
mizchi
@mizchi
jev-playground docs/fit.md. 489 MOBA decisions and 37 chess moves, 0 illegal. A one-line monitor rule is 16/16. Jev on that triage is 6 to 27 wrong pages. The repo root has no LICENSE file.
Mapika
Mapika/decider
decider-4b v2. The v1.4.2 row is 64.13, Intelligence 49.4 against Jev at 53.1. On v1.5 the same row is official rank 5 at 71.3, Intelligence 55.8. The cost cell is an estimate. Apache-2.0 on the board's note.
apolinario
apolinario/decision-index
Decision Index 0.2, dated September 24. Forty benchmarks, chance-corrected. The kit's test locks Jev at 51.67 and AutoJev-27B at 50.94. Code is MIT.
Saman Pandey
@WarBornePhoenix
SamanPandey-in/jevrail, MIT. A hook that scores a shell command before it runs. The printed corpus leaves the model off: 22 of 22 catastrophic commands, and 15 of 26 safe commands still raise an ask.
Lightpanda
@lightpanda_io
A hostel-browsing post with Jev. The linked gist, by katie-lpd, measures browser memory and does not call a model. Default page: 161 MB against Chrome at 1,654 MB.
Tiantong Wu
arXiv:2609.28613
With Wei Yang Bryan Lim, NTU. Decision hijacking on jev-1.13.0. Validated attack success on 510 InjecAgent cases moves from 1.8% to 3.5% when the search can read scores. CC BY 4.0. No repository in the paper text we read.
Wei Yang Bryan Lim
arXiv:2609.28613
With Tiantong Wu, NTU. The same paper. Conditional hit rate at Jev confidence 0.7 stays 0.2% at the final checkpoint. Several registered phrasings do not raise attack success.
Ruoqi Guo
arXiv:2609.29429
First author of RLCDAlignBench, with Yi Liu and Leo Yu Zhang at Griffith and six other coauthors. Generic Noul median AUROC 0.886. Median ECE 0.168. Code MIT. Data CC BY-NC 4.0.
Federico Elia
@FederCr
feder-cr/jev, jevos-1b, built with Loris Salsi. README alt text: 0.815 on 2,000 yes/no questions, against Jev at 0.927. Short latency 54 ms against 344 ms. Choice and Score return 422. pyproject says MIT. No LICENSE file in the root listing.
Roberto Montalti
@rhighs_
rhighs/jev-code. Jev picks grammar productions. The September 18 ladder still fails three coding tasks on commit 517d0f0. Hello world passes in 3 turns and 16 requests. package.json 0.2.0 has no license field.
Louis
@quadrillionboss
Oranquelui/astra-jev-harness, MIT. On one two-file fixture Astra input tokens fall 27.0% and six checks still pass. An older run cut files from 28 to 4 while Astra input fell 1.2%.
Mohit Garg
@mohitt_garg
mohit67890/imajev, Apache-2.0. Imajev-4B is a Qwen3.5-4B LoRA that can read two photos. Changelog v1.4.2.2 ranks it first at 67.37. On the v1.5 addendum it would place sixth at 70.4, outside the ranked order. His September 26 note moves ImajevBench from 82.4% to 83.9%. The v1.5.4 page numbers that 70.4 as official rank 9, still tagged roster addendum. The page opened October 4 is v1.5.6, and imajev-4b is official rank 10 at the same 70.4. Image JevBench v0.1.5 lists 76.4 at rank 2. Opened October 8, 2026, Image JevBench v0.3.0 lists capability 68.0 at rank 1 and composite 63.7 at rank 2 of 52.
Christopher House
crh225/plumb
Apache-2.0 training pipeline on JevK5 v0.2. His public-hard runner, one RTX 4080 Super, is 89 of 111. Changelog v1.4.2.1 scores the submission 65.84. On the v1.5 addendum it would place fourth at 71.6, outside the ranked order. The NOTICE file copyrights the work 2026. No personal X handle on the repo. The v1.5.4 page numbers that 71.6 as official rank 5, still tagged roster addendum.
Sandipan Kundu
@SandipanKundu42
skundu42/kev, a 400M encoder. Not Jared Palmer's Kev. On 17,476 of his own rows the checkpoint scores 79.96% against Jev at 73.54%. On 2,000 PhishNChips emails it flags 1,995 and scores 50.25%. The root listing has no LICENSE file.
David
@dzhng
dzhng/jevgrep, MIT. Jev picks files. Sol still writes the patch. Package 0.4.3 keeps 8 of 10 official solves. Combined cost, Jev's list price included, is 25.8% under a saved baseline. The September 26 post said 40%.
Risk Averse Technology Company
@RiskAverseTech
RiskAverseTech/toolgate, MIT. Version 0.15 scores a shell call after static rules. Set 6 is 19 of 20, with every pair separated. The 1,369-decision week is version 0.9.1, and the friction there was too many asks. Not a sandbox.
Deji
@DejisHQ
A September 28 grants thread. 100 posts from one day, 7 marked a real yes before he saw Jev. After full pages and one post at a time: 75 of 78 nos hidden, 5 of 5 yeses shown. About 2 cents. No repository.
Yusuke
@yusukelp
jevbench.app. Recorded Jev answers on labeled examples. Lead qualification is 20 of 21, and the miss is a yes at 0.51. Opening a test makes no new model call. This site is not Florian S's JevBench.
Christoph Görn
@goern
A Claude Code router. Local jaredpalmer/kev-4b scores 1 of 7 on eight of his prompts. A log of typesafe/jev-router names gpt-6-luna, deepseek-v4.1-flash, and kimi-k3 as the models that answered, with 7 of 16 calls in error.
Violoop
@violoop
Four live desktop runs on September 29. One QuickBooks bill, 123 decisions each: Jev 127.3 s and $0.0027, Claude Opus 5.5 978.9 s and $1.38. The blog byline is BVIO marketing. The page says a local model on Violoop hardware is research, not a current product.
Vi Tran
@viviane_trn
Posted Polylane's write-up with Boris Tane. Five fixed-answer on-call steps moved to jev-1.13.0. The rollup is p90 4,752 ms to 508 ms and $0.76199 to $0.46369 per 1,000 calls. No accuracy.
Boris Tane
@boristane
Coauthor of the Polylane on-call write-up at Coreplane Labs. The page is dated September 25 and the post is September 28.
Tim Lin
@ycalintim
yctimlin/JevScout, MIT. A Claude Code hook that asks Jev which pieces of an oversized MCP result to keep. The post says 38% less agent time and 26% less agent cost across 7 tasks. The printed table has four rows.
Zhaorun Chen
@zrrrr_cn
Posted a DTap chart for Jev 1.13 as a tool selector. Direct misuse attack success is 70.1%, and a second Jev check brings it to 36.1%. The chart does not print a task count.
Sung Kim
@SungKimUp
Posted Solar Decide on OpenRouter. The chart says 87.0% on the public JevBench 231 at a p50 of 0.14 s, against Jev at 86.1% and 0.30 s. p95 is 15.7 s against 0.40 s.
Arnav Gupta
@championswimmer
championswimmer/pi-skill-picker-jev, MIT, 0.1.0. A Pi extension that asks which skill names to load. The README's 20-skill demo cuts about 1,000 tokens from the first prompt.
Tony Gentilcore
@tonygentilcore
Posted Glean's write-up of four internal decisions. On 4,855 queries, Jev Choice cost about $0.00044 and took 0.195 s at p50. Production order as a tie-break raised Recall@6 from 40.2% to 44.3%.
Maincode
@MaincodeAU
Posted Matilda on September 30. The card is a 26.1B model under an Apache 2.0 line, and a self-run of Decision Index 0.2.1 at 59.26 against the board's Jev cell of 57.91. The run section still says the weights are private. Opened October 8, 2026, the card prints Decision Index 0.3 at 60.16, and the model API lists private false.
Dave Lemphers
@davidlem
Replied on the Matilda thread that the release was the top Jev spot, and linked the same Hugging Face repo. That phrase points at the Decision Index self-score, not at JevBench.
Nicholas P. Waltz
PostHog/jeeves
The README citation is Nicholas P. Waltz, 2026, Jeeves: Reasoning Improves Jev-like Decisions. Public JevBench hard, with thinking, is 0.865 against a Kev-published Jev cell of 0.730. No personal X handle on the pages we read.
K.Chatani
@chataclaw
Posted the September 30 pointer at PostHog/jeeves. The post restates the hard-tier pair 0.730 and 0.865 and has no photo. The pages we read do not say Chatani wrote the repository.
kyu1204
kyu1204/jgrep
MIT package, command jgrep, npm name jevgrep. The September 28 bench is mean recall 0.926 on 60 commits at $0.1112. The README says the repository is separate from dzhng/jevgrep.
MINT
@MINTr5dw
Replied with the rounded jgrep figures, 12%, 90%+, and $0.11, on 60 commits. The bench file prints 0.119, 0.926, and $0.1112. The reply does not say MINT wrote the repository.
恒星sun
@vintcessun
Posted Ollaya on September 30. ollaya-dev/ollaya, Apache-2.0, serves POST /v1/systemone on port 11435. The README prints winnow:e4b at 0.722 against TypeSafe Jev at 0.738.
Yubo Li
arXiv:2609.26550
With Yidi Miao, Ramayya Krishnan, and Rema Padman, revised the judge-cascade paper to v3 on September 29. On 4,850 public judgments JEV is 1.7 points behind GPT-6. No personal X handle confirmed on the pages we read.
Cristian Leo
@cristian_le0
Posted a September 28 pointer at the judge paper and quoted an RM-Bench hard/normal pair. We did not find that pair in the v3 sections we opened.
Jianyi Hu
arXiv:2609.31142
Listed first on the JevAdvBench abstract, submitted September 25, with Hangtao Zhang and six coauthors. The abstract says one appended opinion flips 12.1% of jev-1.13.0 decisions. No personal X handle confirmed on the pages we read.
Isao Takaesu
@bbr_bbq
Posted the September 28 pointer at JevAdvBench. The post has no photo we saved. We read the abstract, not the 19 tables.
OpenAI Developers
@OpenAIDevs
Posted the Decisions API limited preview on September 29, powered by GPT-6 Luna, for text or images. On October 6 at 20:47 UTC the account posted public beta for all developers. On October 22 at 19:07 UTC it posted general availability. Date headers 02:23:12 GMT and 02:23:13 GMT October 23, two POSTs from this desk returned HTTP 200.
Tibo Sottiaux
@thsottiaux
Posted the Decisions API on September 29 at 17:27 UTC, before the OpenAI Developers announcement. Bio line: Codex & ChatGPT @OpenAI. The post says it supports visual inputs and is tuned for less than a few hundreds of milliseconds end to end. On October 6 at 20:56 UTC a day-two roundup says Decisions API live for builders.
chaliy
everruns
Merged everruns pull 3924 on September 30. That pull request's live smoke against POST /v1/decisions returned HTTP 403, Decision API is not enabled for this user, and it says the request shape is unverified. On October 6 this desk's own POST returned HTTP 200. The guide is published. We did not rerun the pull request.
Josh Brown
@joshuafbrown
Posted the Perch Bug Bench chart on September 29. Average precision on 2,220 public before-and-after pairs is 46.8% for DiffusionGemma Jev and 46.2% for Jev. The parent post says the rule-based checks were the strongest result. perchscan.com/benchmark-results returned 404.
Celeris
@Celeris_ai
Posted Celeris-1 Decision on October 8, 2026, at 23:34 UTC. The blog prints 80.0% on 3,210 jev-bench items at a 67 ms median, $0.04 per million input tokens. Opened October 16, Artificial Analysis Decision Index lists it at 52. GET /v1/models with no key returned HTTP 401.
Tom Hamer
@tom_w_hamer
Byline on the Celeris-1 Decision blog, dated October 8, 2026. Quoted the launch at 23:37 UTC and printed 80% on Jev-Bench, 67 ms, and a one-sentence explanation on request.
Liquid AI
@liquidai
Posted a September 29 chart captioned as an internal reproduction of Decision Index 0.2.1. d1 is 58.9 and Jev 1.13 is 57.9. The docs name model d1:free on api.liquid.ai. On October 7, 2026 the account posted Open d1, open weights d1-3B and d1-omni-600M. The d1-3B card prints Decision Index 0.2.1 at 48.57, self-scored. That figure is a different test from the September 29 chart. On October 15, 17:00 UTC, the account posted d1-ultra. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57. That day's Artificial Analysis thread prints an 85% non-hallucination rate on AA-Omniscience-Decision against 2% for hosted d1.
Aurélien Lac
@Aurelien_L_
Liquid AI research engineer. Quoted the d1-ultra launch on October 15, 2026, at 17:14 UTC, and called it the lab's strongest decision model to date.
Respan
@RespanAI
Posted Span-02 on October 15, 2026, 17:00 UTC. On October 16, at 16:01 UTC, the account restated $0.01 per million input tokens and posted a 0-12 chess clip against Jev, reasoning off. The blog prints Decision Index 0.3 thinking off at 61.2 against Jev 1.13.0 at 58.0. The Hugging Face card prints text index 63.62. This desk did not run the weights.
Frank Chen
@francchen
Byline on Respan's Span-02 blog, dated October 15, 2026.
Harvey Li
@hli0407
Posted on October 15, 2026, at 20:21 UTC, that Span-02's engine is up to 7 times faster than vLLM on a single RTX 5090. On October 16, at 22:08 UTC, he wrote that Span-02 is +7.8, 2 times faster and 2 times cheaper than pplx-decider-flash-v1.2 on Decision Index.
Hendrix Liu
@HendrixLiu29208
Respan cofounder. Posted October 16, 2026, 23:16 UTC, that Span-02 had more usage in under 24 hours than Span-01 did in 20 days. The post does not print a request count.
Federica
@federica_freddi
Sqwish Labs cofounder. Posted October 16, 2026, 22:30 UTC, that kettei-fat is #1 of 37 on JevBench, intelligence 80 against 70, and #1 of 60 on JevImageBench. Opened October 17, the API board ranks 39 and prints composite 77.8, Intelligence 79.8.
nish
@Firebird15728
Sqwish Labs. Posted October 16, 2026, 22:41 UTC, that kettei-fat is #1 on JevBench, chub is #4, and #1 on ImageJevBench. The chart on that post is the share image on the kettei-fat page.
Sqwish Labs
@sqwishai
Posted a kettei chub Super Mario Bros clip on October 16, 2026, 10:29 UTC. The blog dated October 14 names kettei fat, chub, core, and dot. Docs price kettei fat at $0.06 per million input tokens. POST https://api.sqwish.ai/v1/decide.
Amelia Charles
vercel.com/blog
Byline with Harpreet Arora on Vercel's October 2026 AI Gateway Production Index, published October 15, 2026. The launch table prints Jev at 35.3% of paid teams in its first two weeks.
Russell Brandom
techcrunch.com
TechCrunch, October 16, 2026, 15:33 UTC. Cites the October index and rounds Jev's two-week paid-team reach to 35%, nearly double Sonnet 5.5.
Nace AI
@NaceAI
Posted Drex Max on October 19, 2026, 13:33 UTC. The product page prints Decision Index 0.3.2 public 67.59 against Jev 1.14 at 60.70. Docs price input at $0.06 per million tokens. Posted Drex 1.5 on September 30 and called it first on Decision Index 0.2.1 and on JevBench. The chart on the next post prints 199 of 231 against Jev at 201 of 231. On October 6, 2026 the account posted open weights and called them Drex 1.1. The card title is Drex DLM. That card's Decision Index 0.2 cell is 52.31 against the kit's Jev cell of 51.67. On October 9 the account posted Drex v1.5 weights. That card's Decision Index 0.3.1 public index is 58.08 against Jev 1.13.0 at 57.96.
Levanto Labs
@levantolabs
Builds Sage, a hosted decision API. Opened October 17, 2026, the JevBench API board ranks Sage 1.3.0 third at composite 74.0, behind kettei-fat at 77.8. Opened October 11, that board ranked Sage first of 27 at 74.0. The pricing page lists Pro input at $0.042 per million tokens and output at $10 per million for reasoning. Big Iron Chris, at the same lab, posted the SageCraft play log on October 9.
Jonathan McKinney
H2O.ai
Wrote the October 7, 2026 H2O blog. It prints H2O-Lightning-4B v1.1 at JevBench composite 72.5 against Jev 1.13.0 at 71.5, on board revision v1.7.17.
Stanislav Sorokin
@stas_sorokin_
Posted OpenBles Firewall 1 on October 9, 2026. The README scores 119 of 120 agent attacks caught, with none of 120 benign look-alikes flagged. The adapter is MIT, on fastino/GLiNER2.5-Decide.
Mani
@mirzaei_mani
Posted Mizan 4B on October 11, 2026, 13:50 UTC. The card xmanii/Mizan-4B prints a mean of 78.7 on 3,000 real Persian records, against a Jev cell of 83.7. Apache-2.0, from Qwen3.5-4B. The post says image input comes in a later update.
doofz
@reversedoofz
Posted THX-01 on October 11, 2026, 10:23 UTC. The card averages 98.4 on 2,843 support tickets, against Jev 1.13 at 97.4. The citation names Farid Aghayev and Elturan Ahmadbayli at HAL-X AI. Apache 2.0, 322M, on mmBERT-base. The post calls the API free. The docs page we opened does not print a price.
Chinh Nguyen
chinhnc
Released Quyet 1.0. The homepage, opened October 12, 2026, prints frozen JevBench v1.6.0 at official score 71.39 and Capability 81.7 for Quyet-1.0-Large. The Large card says Apache-2.0, 31.3B, from Gemma 4 31B.
Pixilab
PixilabAI
Published Blink v0.3 on October 6, 2026. The card prints a self-run of Decision Index 0.2.1 at 57.48, against the board's Jev cell of 57.91. The license field is Apache-2.0. The base is Gemma 4 26B-A4B.
Rox
@rox_ai
Posted a September 28 reranker chart, mean over 75 queries. At 200 chunks Jev-Score is 96% and production, labelled GPT-5 Mini, is 84%.
Shapor Naghibzadeh
QueryStory
Built the Jev Sentinel demo named in Tim Fernholz's September 30 TechCrunch piece. The demo page prints about $55 per million actions. TechCrunch prints $2.94 against $372 for one job.
Inception
@_inception_ai
Posted Mercury Decide on September 30 and called it the most intelligent decision model on OpenRouter, citing JevBench v1.4, at up to 14 decisions a second. The post prints no score. Opened October 8, 2026, the JevBench API board ranks it third at composite 72.4. The free model page returned the same day, priced Free.
Tim Fernholz
TechCrunch
Wrote the September 30 piece on OpenAI's Decisions API and on Jev Sentinel. He also wrote the September 18 interview already on this desk.
Olaverse
@olaverse_
Posted PurpleMIST-Flash-1.0 on October 10, 2026. The card scores 0.668 accuracy on 15,050 translated African-language decisions. The account does not print a personal name on the card we opened.
Cloudflare
@Cloudflare
Posted Clef and Clef-flash on October 1, 2026 at 19:51 UTC. The post calls them open-source decision models on Workers AI and links the Birthday Week blog. The post has a link card and no attached photo. On October 9 at 23:11 UTC the account posted Clef-omni. That post has no attached still.
Michelle Chen
@michellechen
Posted Clairic on October 19, 2026, 17:54 UTC, a System One extraction model on Workers AI. The blog cascade table is 19.3 s against 61.7 s. Byline on the October 1 Clef blog. On September 18 she posted a DiffusionGemma demo that scores token probabilities on Workers AI, at kev.workers-ai-mle.workers.dev. The Clef post says the later models use a Qwen backbone. She is also the byline on the October 9 Clef-omni post.
Strands Labs
strandsagents.com
Posted Strands Decider 2B on October 1, 2026. The checkpoint is Apache 2.0, 1.9 billion parameters. v19 scores 167 of 231 on the public JevBench tasks. The blog does not name an author.
Tianwei Mu
arXiv:2610.02048
Listed first on HydroJEV, submitted October 1, with Shengyan Jiang and six coauthors. jev-1.13.0 is the screen on simulated EPANET windows. Mean in-distribution macro-F1 is 0.63. No personal X handle confirmed on the pages we read.
Hekate
@iCleanAI
Posted HydroJEV on October 2, status 2105910123596280015. The post restates the spare-the-reviewer line and says the evaluations are simulated. The post has no attached photo.
Yiqun Zhang
arXiv:2609.35293
Listed first on the DimABSA paper, submitted September 28, with Peidong Wang, Zihan Wang, and Shi Feng. Frozen jev-1.13.0 plus 488 coefficients. Task 1 RMSE is 1.0645. No personal X handle confirmed on the pages we read.
PageIndex
@PageIndexAI
Posted VectifyAI/jev-doc-search on October 2, 2026, 05:44 UTC. The repository asks jev-1.13.0 to walk a PageIndex tree. The printed table is two filings, both marked correct.
Erik Spock Gafni
@EGafni
Quoted the PageIndex post on October 2, status 2105900570800496742. The sentence adds no measurement.
John Crown
@johncrown
Co-founder of Wrangle. Posted the WrangleBench chart on September 30, 2026, 19:46 UTC. With Jev the bar is 137.3 perfect matches per 200. Without Jev it is 105.1.
Mikołaj Mizera
@AjMizera
Posted the MMAI Gym chart on October 2, 2026, 09:12 UTC, and tagged Insilico Medicine. Five CYP assays, 12,563 molecules. Jev 1.13's calibration curves stop at 0.69.
Edward Irby
@edwardirby
Engineering at You.com. Posted the risk-sweep write-up on October 1, 2026, 16:25 UTC. Three escalated sweeps total $0.43. The repository is MIT.
Bud Ecosystem
@BudEcosystem
Posted Bud Decision Studio on October 2, 2026, 08:42 UTC. The app serves eleven open decision models at 127.0.0.1:8420. GitHub lists no license on the repository.
Drew Breunig
@dbreunig
Posted cmpnd-ai/dspy-system-one-agent-patterns on October 1, 2026. config.py pins jev-1.13.0. The shell gate refuses below 0.2 and runs at 0.6 or above. He is also named on the DSPy 3.4.0 Jev notes.
Isaac Miller
github.com/isaacbmiller
Released DSPy 3.4.0 on September 25, 2026. The notes add the TypeSafe client and the experimental Noul, Choice, and Score types. No personal X handle confirmed on the pages we read.
Nathan LeClaire
@dotpem
Posted office-hours slides on September 30, 2026, 22:01 UTC. The cost chart prints Luna at 92.5 percent and $0.746 per 1,000, and a live threshold near 0.78 at 88.0 percent and $0.583 per 1,000. On October 9, 2026 he posted that Jev is a reranker, and not just for text. That post has no table. On October 12 he posted that pairwise Choice plus Bradley-Terry seemed better than ranking by a Score. On October 15, at 20:10 UTC, he posted that Jev's new Box primitive is Yolo for everything.
Nick Gideo
@NickGideo
Author of question-kit, MIT, 0.2.0. On October 12, 2026 he committed rankByScore and rankPairwise on branch order-v2-and-core-updates, commit 25a3b71, still listed as Unreleased, and wrote that he would publish tomorrow.
Didier Lopes
@didier_lopes
October 12, 2026 quote of LeClaire's ranking tip. He wrote that he had used pairwise Jev for an Elo leaderboard of tickers. The quote has no table.
Tobias Petry
@tobias_petry
October 12, 2026. Asked for a TypeSafe ranking primitive: n values in ranked order, 1.0 for the winner, and a relative score in [0, 1]. The primitives pages that day still listed Choice, Score, and Noul.
Hamed Nilforoshan
@h_nilforoshan
Posted a HiringCafe thread on October 7, 2026. Query-to-job Spearman is 0.74 for Jev and 0.71 for OpenAI. Resume-to-job is 0.71 and 0.67. Gemini 3.1 FL is 0.63 and 0.61. The thread does not print the pair count. It says HiringCafe has 2.5 million users.
Secretof_forest
@Secretof_forest
Posted the PolicyLM-1.7B card on October 7, 2026, 03:49 UTC. The numbers on this desk come from the card. We did not find an official Musubi Labs post.
Musubi Labs
huggingface.co/musubilabs
The PolicyLM-1.7B card, opened October 7, 2026, prints custom-policy accuracy 0.842 at a 0.5 cutoff. The model returns a score from 0 to 1 for each category, up to 16 in one pass. No X handle for the org was on the pages we read.
Perplexity
@perplexitydevs
Posted pplx-decider-v1.2 and pplx-decider-flash-v1.2 on October 16, 2026, 18:42 UTC. The v1.2 card prints Decision Index 0.3 overall 68.09 on joint requests against v1.1 at 62.30. Docs price v1.2 at $0.02 per million input tokens and flash at $0.01. Posted pplx-decider-v1-27b on October 1, 18:24 UTC. The chart is a 7,210-row panel, 85.71% against jev-1.13.0 at 84.51%. On October 6 the account posted v1.1 at $0.02. Decision Index 0.3, opened October 7, ranked v1.1 first at 62.8.
David Hendrickson
@TeksEdge
On October 3, 2026, 06:00 UTC, reported a ramgpt EXL3 build of pplx-decider-v1-27b at 4.00 bits per weight, 15.35 GB, medians 135 ms, 278 ms, and 430 ms. He says that card has no BF16 accuracy comparison. The link was in the image alt text.
ABE Masanori
@abe_masanori
Posted a Qiita comparison of Snowflake AI_FILTER and Jev Noul on October 24, 2026, 11:20 UTC. 1,774 of his 2025 X posts. Jev about $0.055 against AI_FILTER about $1.87. AUC 0.9685. Run 2026-10-20, response jev-1.13.0.
Qi Li
arxiv.org/abs/2610.18773
First author of Jeva, with Xinchao Wang, National University of Singapore. DailyPapers posted the paper on October 14, 2026, 20:10 UTC. Hugging Face LIQIIIII/Jeva-4B. This desk did not find a personal X handle on the pages it opened.
Xinchao Wang
liqiiiii.github.io/Jeva
Corresponding author on Jeva. The paper prints 0 flips on 600 Choice items and BANKING77 at K=16 91.62 against Jev-1.13.0 at 82.50.
George Maloney
@george_onx
Posted GLiDE on October 1, 2026, 18:27 UTC. The text says 6.9 points over Jev on the Decision Index and 11.5 points on Knowledge and Reasoning. The blog's self-run is 64.81 against 57.91. On October 22, 16:36 UTC, he posted GLiDE v1.2 at 71.5 Full, 82 ms, 1.9% ECE. On October 23 he printed 11.3 points, 93 ms median, ACOS 27.2 to 61.6, and non-hallucination 16 to about 78.
Sahibzada Allahyar
@singularity_bly
Byline on the GLiDE blog with Mary Newhauser. On October 3 he posted the bar chart: GLiDE 64.81, Clef 61.2, Jev 57.91, Clef-flash 57.1, pplx-decider 56.39. The axis starts at 56. The v1.2 training blog, byline October 22, is his. It prints Decision Index 0.4.1 Full 71.5.
Mary Newhauser
fastino.ai
Named with Sahibzada Allahyar on the October 1 GLiDE blog. The page prints the 64.81 self-run. No personal X handle confirmed on the pages we read.
AutoTrust AI
@AutoTrustAI
Posted GEV-26B-Decide on October 3, 2026, 05:29 UTC. The card prints an adaptive Decision Index self-score of 62.48. The posts are video. No personal author name on the card we read. The October 4 follow-up is JEV-27B-VL. The card's October 3 section prints 15 of 20 pick-and-place scenes.
Jacob Lin
@JacobLinCool
Posted StreamDecisionBench on October 2, 2026, 17:41 UTC. The chart prints Kev-27B at 62.15 and the Jev API at 59.63. The live page HTML did not render the table.
Jhen-Ke Lin
arXiv:2609.38612
Listed first on StreamDecisionBench, with Chung Chun Wang, NYCU. Submitted September 29, version 2 dated October 1. The paper table is a different scoring from the October 2 chart. The pages we read do not say this is the same person as @JacobLinCool.
Chung Chun Wang
arXiv:2609.38612
Second author on arXiv:2609.38612, with Jhen-Ke Lin. No personal X handle confirmed on the pages we read.
Pranjal Aggarwal
@PranjalAggarw16
Posted the CUA Speedrun chart on October 2, 2026, 18:36 UTC. The callout is Jev at 9.8% and 125 seconds a task on 50 OSWorld tasks. The code tree is cua-speedrun/cua-speedrun, branch jev-agent.
Tobias Deußer
arXiv:2609.37647
Listed first on the September 29 abstract, with Lorenz Sparrenberg and Rafet Sifa. 346,009 requests, under $10, Jev ahead of Qwen3.8-27B on 27 of 37 datasets. We read the abstract, not the PDF.
Databricks
@databricks
Posted on October 16, 2026, 19:17 UTC, that TypeSafe's Jev-1.14 is available through ai_classify and ai_decide. Posted ai_decide on October 1, 13:54 UTC. The September 30 blog byline is Jane Zhang, Matthew Ding, and Archika Dogra. Beta. The docs do not print a price.
Gautam Kishore
@gautamkishore
Posted NIRNAY 450m on October 3, 2026. Apache-2.0 Laya fine-tune. phase_b is 0.8792 on 3,080 Banking77 rows. The post's 0.8656 is phase_a. He says the Jev cell was not measured in the repo.
Samrat Dutta
@TheSamratDutta
Posted WaterSheep on October 3, 2026, as a reply. Apache-2.0 ModernBERT-base fine-tune. POST /v1/systemone on port 8766. 77.8% in distribution, 61.2% on held-out datasets, no Jev column.
Omar Hernandez
ohernandezdev
jevmod 0.2.1, MIT. One Noul per category. Adult text flags at 0.80. Sexual AUROC 0.984 on OpenAI's moderation eval, remeasured September 23. No X handle confirmed on the pages we read.
SeYeD.Dev
@seyed_dev
Posted ZeroNSFWBot on September 22, 2026. GitHub mohamadkhoshnava/ZeroNSFWBot. Pictures are scored on the host. Jev reads clipped text when a key is set. The README prints no accuracy on adult images.
StartLux
@StartLuxAI
Posted StartLux-Decision on October 3, 2026, 13:41 UTC. The README, opened October 20, 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. No person is named on the post or the README.
Remek Kinas
@KinasRemek
Posted basal-omni on October 17, 2026, 14:15 UTC. Opened October 19, JevBench ranks basal-omni-1.0-ultra first of 184 open systems on Capability at 82.5. The card prints Werdykt v1 at 0.879. Apache-2.0, Gemma 4 31B. Posted basal-1.5 on October 4, 06:58 UTC. Werdykt v1 puts the 4.5B at 72.12% and Jev 1.13.0 at 81.64%.
Vyom Patel
@vyomnpatel
Founder of Teenage Works. Posted ToddlerJev on October 18, 2026, 22:19 UTC. MMAU-mini 68.70%. Weights TeenageWorks/Toddler-Jev under the Qwen Research License. This desk did not run them.
Kenneth Yeung and Kasimir Schulz
HiddenLayer Research
Published Jevil in the details on October 14, 2026. TypeSafe Jev API on a checkout ad: Confirm Price 91 percent, Place order 9 percent. Thirty toxic prompts flagged at baseline, then TokenBreak edits. Date headers 07:24:57 GMT and 07:24:58 GMT October 19, this desk's jev-1.13.0 still chose place_order, 1.0 then 0.89.
Nick Frichette and Feras Al-Kassar
@Frichette_n
Datadog Security Labs. Posted October 19, 2026, 19:12 UTC. Article dated October 15. Jev Noul recall 82.90% on raw attacks, 3.88% with a counterfeit policy prepended. LangChain is_risky 40-run mean 0.893 then 0.290. Date headers 08:50:46 GMT and 08:50:47 GMT October 20, this desk's jev-1.13.0 is_risky was 0.94 then 0.13.
Kevin Flansburg
@kevin_flansburg
Workers AI at Cloudflare. Posted Clairic on October 19, 2026, 18:24 UTC, as a system one model based on regression instead of classification, and said open weights are coming soon.
Cuth
@ItsCuthulhu
Posted Sieve 31B on October 20, 2026, 01:50 UTC. The text prints 63.8 on the index's full suite, second globally, and a LoRA adapter on Gemma 4 31B. The Hugging Face card prints Full score 63.8. sthanika.ai prints tied 3rd of 174.
Sridhar Muppidi
PurpleTalk
Co-founder of PurpleTalk. A GlobeNewswire copy dated October 19, 2026 quotes him on Sieve 31B. The Hugging Face card prints Decision Index 0.3.2 Full score 63.8. This desk did not run the weights.
Ying Fang, Zhifu Gao, Qing Gu, Jiajun He, Kangqi Jing, Xiangang Li, Yu Pu, Wupeng Wang
arXiv 2610.21614
Qwen-Audio-Jev, submitted October 15, 2026. Choice 78.70% on MMAU-mini, about 20x against Qwen-Audio-3.1-ASR-Flash-Next. 1LittleCoder posted the abstract on October 18, 21:14 UTC. This desk did not run the model.
Xunzhuo Liu
@XunzhuoLiu
Posted Decision 2.0 on October 3, 2026, 01:25 UTC. Six Apache-2.0 checkpoints. Vega's card prints 56.5 on its own Decision Index 0.2.1 run and does not print a Jev row. The vLLM account posted the same release two minutes earlier. On October 7, 2026, 07:04 UTC, Liu posted Vela 2.0. The vLLM account had posted that release at 04:30 UTC. The 9B card prints 41.63 on Decision Index 0.2.1 with Noul calibration on. On October 11, 2026, 08:59 UTC, Liu posted Decision 3.0. The d3 card that day printed 64.7 on its Decision Index 0.3.1 table. Opened October 16, the same card prints 0.3.2 at 66.3. At 11:35 UTC he posted a Box demo on d3 and d3-flash. On October 20, 00:42 UTC, he posted d3-turbo. On October 27, 03:01 UTC, he posted d3-ultra. Opened October 26, Decision Index 0.4.1 Full score is 70.18 for d3-ultra and 67.31 for d3-turbo.
Yuma Ichikawa
@yuma_1_or
Posted Decision Model Factory on October 27, 2026, at 03:19 UTC, and linked arXiv 2610.34956. Table 1 prints DF-JA-Large at 67.63 on S1MB Japanese Task Avg against a fixed public snapshot high of 60.05. A Hugging Face search that day returned no matching public checkpoint.
Together AI
@togethercompute
Posted tev1-4B-experimental on September 23, 2026, 22:06 UTC, at $0.042 per million input tokens and $0 output. The model page opened October 7, 2026 lists $0.04 and free output. The chat call returns one letter. Ollama serves tev1 on /v1/systemone.
ForkScape
@forkscape
Posted fsdecide on October 5, 2026, 03:26 UTC. The GitHub account on the repository is dreamerron. Review is 91.3% of 80 against jev-latest at 56.3%. On the authors' general set, jev-latest is 90.5% and fsdecide is 60.5%. The README says MIT.
Zachi
@iam_zachi
realZachi/pg-jev, PostgreSQL License. WHERE jev(table, condition) on PostgreSQL 14 through 17. The README prints about 3.5 seconds and $0.012 on 2,000 rows, then about 50 ms from cache. Batches of 20 were 100% on his ground-truth check. pgjev.com still prints batches of 40.
Avi Chawla
@_avichawla
October 7 post and the September 30 newsletter describe a Jev stage after hybrid retrieval. The printed passage scores, C1 0.93 through C4 0.09, are an illustration. The threshold in the example is 0.70.
Ben Bartholomew
Hindsight
September 24 note on the typesafe reranker in Hindsight 0.10.1. Listwise Choice on 30 LoCoMo candidates: recall@1 0.950 against MiniLM at 0.800. Pruning is off by default and drops 19% of the gold.
Nicolò Boschi
@nicoloboschi
October 2 post. Hindsight's Jev reranker against bge-reranker-v2-m3 on NFCorpus, 323 queries and 11,139 memories. The post prints gaps of +4.4% nDCG@1, +5.5% MRR, and +8.3% Hit@10, and does not print the absolute cells.
Akshat Kumar
lyzr.ai
October 6. 21,314 calls to jev-1.13.0, $0.72. One-call rerank, SciFact NDCG@10 0.836 against dense retrieval at 0.776. Sufficiency AUROC about 0.90 on SQuAD v2 and on HotpotQA. No personal X handle confirmed on the page we read.
Jayanth Penumarthi
Thoughtworks
October 6 design note. Bands above 0.9, 0.6 to 0.9, and below 0.6. The note prints no table. No personal X handle confirmed on the page we read.
Arena
@arena
Posted an Agent Arena run of typesafe/jev-router on October 7, 2026, 22:29 UTC. More than 4,700 sessions. The thread prints +3.6% net and a median $0.155 a task, and says a direct DeepSeek V4.1 Flash call is cheaper and faster at similar performance.
Doug Turnbull
@softwaredoug
Posted the bag-of-decisions note on October 8, 2026. On the WANDS CSV, one Noul has mean NDCG@10 0.574262 and a GPT-5 question list has 0.609814. Both medians are 0.5609447680982702. The score adds 10 times the summed Nouls to retrieval.
Ravi
@ravikml
Replied on October 9, 2026 that a question other than relevance worked better, and linked Doug Turnbull's note. The bio on that post says AI/ML Engineering at ClickHouse.
laxman
@llmluthor
Posted Step-Jev on October 8, 2026, 17:20 UTC. A Qwen3-4B search agent on BrowseComp-Plus scores 19.6% when the reward grades the final answer and every step, and 5.4% when it grades only the final answer. The code is bandr-ai/bandits. The GitHub license field was empty on October 9.
Achint Srivastava
Microsoft, Office of the CTO
Posted Microsoft-Decision-1 on Command Line on October 9, 2026. Input is $0.042 per million tokens and output is free. The post says the model was 4.5 times quicker than Quyet-1.0-Large and 35 times quicker than GPT-6 Sol, and it does not print an accuracy. The byline is VP of Software Engineering.
Satya Nadella
@satyanadella
Posted the Microsoft-Decision-1 announcement on October 9, 2026, 18:37 UTC. The attachment is a 10.2 second video. We did not transcribe a score from it.
OpenRouter
@OpenRouter
Posted Armor in beta on October 15, 2026, 21:00 UTC. The 492-call chart prints Jev 1.13 at 90.7%, 446 of 492, $0.0021 a call. Posted on October 15, 00:16 UTC, that Jev 1.14 is now on OpenRouter. Opened October 16, typesafe/jev-1.14 lists $0.042 per million input tokens, output free, a 262K context, and TypeSafe P50 0.19 s. ~typesafe/jev-latest already linked Jev 1.14. GET of that id still returned HTTP 404. Posted on October 9, 2026, 22:43 UTC, that Microsoft-Decision-1 was live. The post quotes $0.042 per million input tokens, free output, and a 32K context. The model page, opened October 10, lists Azure at a P50 of 0.18 seconds.
Alex Atallah
@alexatallah
OpenRouter CEO. Posted on October 15, 2026, at 23:42 UTC, that Armor is an auto permissions mode powered by decision models like Jev.
Yuuki Yamashita
@yama3133
Posted a Qiita write-up dated October 15, 2026, scoring jev-1.14.0 at 0.833 and jev-1.13.0 at 0.817 on 60 Kyoto-ben tickets, three-run means. Majority vote matched on all 60. McNemar p = 1.000. This desk's Choice on the published warning phrase chose social_nicety on both models, n = 3.
Cristian
@XCSme
Posted on October 15, 2026, at 12:21 UTC, that Jev 1.14 seems to do worse than 1.13 on his tests. AI BENCHY, opened later that UTC day, prints Jev 1.13 default at 2.6, 6 of 12 tests correct, and Jev 1.14 default at 2.3, 4 of 12. Adapters supply choices.
Egor Merkushev
@egormerkushev
Posted on October 15, 2026, at 13:59 UTC, Released Jev4Mellea 0.7.0 with experimental image support. The October 15 visual comparison records twelve scenes on jev-1.14.0, coverage 5 of 12 at 0.10 / 0.90. This desk did not install the package.
brainstormity
@brainstormity
Posted on October 15, 2026, at 14:08 UTC, open-sourced JEVsnap, a screenshot redactor on Jev 1.14. github.com/brainstormity/JEVsnap is MIT, v0.2. Default AI path sends OCR text. Box visual check is off by default. This desk did not install the app.
Göksu Güneş
@seoilan
Posted on October 15, 2026, at 10:02 UTC, that both Jev 1.13 and Jev 1.14 scored 29 of 30 on SosyalDijital's 30 synthetic Turkish tickets, and that both marked the same buyer as a seller. The article records OpenRouter model fields typesafe/jev-1.13-20260917 and typesafe/jev-1.14-20261014, and a cost of $0.001129884 for 60 calls.
TypeLLM
@TypeLLM
Posted on October 16, 2026, at 05:44 UTC, that Jev 1.14 cannot return an array of vehicle boxes, including Tesla's flying Roadster. The cookbook at 14.00 s prints six boxes. This desk's every-vehicle Box returned one rectangle, n = 3, and type array returned HTTP 400. Posted on October 15 at 05:47 UTC that TypeLLM now supports image bounding boxes. The find-objects example prints a dog_head box at 0.352 / 0.652 / 0.344 / 0.628. The September 23 file scores Qwen3.8-27B at 228 of 231 public JevBench tasks with thinking. The homepage, opened October 10, prices input at $0.05 per million tokens.
Alexander Whedon
@alex_whedon
Posted SubQ Snap on October 12, 2026, 16:03 UTC. Posted Snap 1.1 on October 22, 19:48 UTC, at $0.04 per million input tokens, When2Call 96.2, and a tie at 65 on Artificial Analysis Decision Index. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs from this desk returned HTTP 200.
Subquadratic
@subquadratic
Quoted the SubQ Snap launch at 16:05 UTC on October 12, 2026, and Snap 1.1 at 20:05 UTC on October 22. Docs opened October 23 name subq-snap-1.1 at $0.04 per million input tokens. Date headers 02:23:17 GMT to 02:23:20 GMT October 23, three POSTs from this desk returned HTTP 200.
Base Labs
@baselabs
Posted mammoth-1 and elephant-1 on October 21, 2026, 17:14 UTC. The card prints Decision Index 0.4 Full 71.24, first of 208. Opened October 23, live 0.4.1 Full 68.4 rank 3. MIT, from Qwen3.5-27B and Qwen3.5-9B. Date headers 02:23:21 GMT to 02:23:23 GMT, GET /v1/models with no key returned HTTP 401.
Baseten
@baseten
Posted mammoth-1 and elephant-1 as free for one week on Model APIs, October 21, 2026, 17:13 UTC. The free week ends October 27, 11:59 PM PT. POST https://inference.baseten.co/v1/systemone.
Vals AI
@ValsAI
Posted on October 22, 2026, 23:57 UTC that d1-ultra is the only decision model to get all 1,960 ClaimBench claims right. The October 20 table ranked Decider V1.2 first among decision models at 99.69%.