Updated
Desk
Repos
GitHub projects this digest has already cited. A week-one roundup lists more catalogs than this page repeats. For the full URL list, see sources.
Local stand-ins
alphaXiv/OR-decision-4B
Apache-2.0 LoRA on Qwen3.5-4B. Card JevBench primary 193 of 231, 83.55%. Launch chart Choice 79.7 against Jev 79.6, sarcasm 53.6 against Jev 70.7. Corpus 117,186 train examples. OpenResearch at github.com/alphaXiv/OpenResearch. This desk did not run the weights.
ytbai/MiniCPM5-2B-Jev
Apache-2.0 on openbmb/MiniCPM5-2B. Project prints JevBench 78.79%, first among open models at or under 2B, and decision-v7 86.41% with ECE 0.0307. Code github.com/yuting-ai/minicpm5-2b-jev. Ollama ActBro/minicpm5-2b-jev. This desk did not run the weights.
ndemirdemir/BranchOne
Apache-2.0. Qwen3.5-0.8B-Base plus LoRA and a shared scoring head. Returns one probability per supplied candidate. Author A100 Colab median 0.305 s on ten short text questions. Code github.com/ndemir/BranchOne. No published JevBench score. This desk did not run the weights.
hyperlogue/decisive
Rick Ye's WebGPU decision-model runtime. Demo pplx-decider-flash. Wrapper and local runtime, not a catalog model. This desk did not run the demo.
fastino/GLiDE
GLiDE v1.2 weights. Fastino Community License, modified Qwen3.8-27B. Card Decision Index 0.4.1 Full 71.5, rank 1 of 229, opened October 26. POST https://api.fastino.ai/v1/systemone, model fastino/GLiDE. Date headers 08:49:09 GMT to 08:49:11 GMT, GET /v1/models omitted GLiDE, POST /v1/systemone HTTP 401. This desk did not run the weights.
LIQIIIII/Jeva-4B
Apache-2.0, Qwen3.5-4B. Paper 0 flips on 600 Choice items, BANKING77 K=16 91.62 against Jev-1.13.0 at 82.50. Code github.com/LiQiiiii/Jeva. This desk did not run the weights.
baselabs/mammoth-1
MIT, Qwen3.5-27B. Card Decision Index 0.4 Full 71.24. Live 0.4.1 Full 68.4 rank 3, opened October 23. Sibling baselabs/elephant-1. POST https://inference.baseten.co/v1/systemone. Date headers 02:23:21 GMT to 02:23:23 GMT, GET /v1/models with no key returned HTTP 401. This desk did not run the weights.
SubQ Snap 1.1 docs
Model id subq-snap-1.1, $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.
sthanika-ai/Sieve-31B
Apache-2.0 LoRA on google/gemma-4-31B-it. Decision Index 0.3.2 Full score 63.8, opened October 20. Authors' public-suite run 62.52. Code github.com/sthanika-ai/Sieve. This desk did not run the weights.
Drex Max docs
POST https://console.nace.ai/v1/systemone, model drex-max. $0.06 per million input tokens. Product page Decision Index 0.3.2 public 67.59. Date headers 08:50:01 GMT and 08:50:02 GMT October 20, GET /v1/models with no key returned HTTP 401. This desk did not send a billed POST.
Workers AI, clairic
Preview @cf/cloudflare/clairic. $0.10 per million input tokens, 65,536-token window, vision. Extraction model, not Choice/Score/Noul. This desk did not call the API.
LangChain TypeSafe tool-risk gating
Experimental Noul is_risky, block at 0.5. Datadog's 40-run mean on a viewer who cannot edit is 0.893, then 0.290 with a counterfeit definition in profile_description. This desk did not install langchain-typesafe.
Remek/basal-omni-1.0-ultra
Apache 2.0, Gemma 4 31B. POST /v1/systemone on basal-serve. Opened October 19, JevBench Capability 82.5, first of 184 open systems. The card prints Werdykt v1 at 0.879. This desk did not run the weights.
TeenageWorks/Toddler-Jev
Qwen Research License. Base Qwen2.5-Omni-3B. MMAU-mini 68.70%. Posted October 18, 2026. This desk did not run the weights.
browser-use/jev-ultrafast
HiddenLayer ran a checkout through this harness. TypeSafe Jev API: Confirm Price 91 percent, Place order 9 percent after an ad. This desk did not rerun that harness.
perplexity-ai/pplx-decider-v1.2
Posted October 16, 2026. Fine-tune of Qwen3.8-27B. Card Decision Index 0.3 overall 68.09 on joint requests against v1.1 at 62.30. Model API lists apache-2.0, private false, gated false. The README usage section still says the repository is private. Flash is perplexity-ai/pplx-decider-flash-v1.2. This desk did not load the weights.
Sqwish OneHot
POST https://api.sqwish.ai/v1/decide. Model ids kettei-fat, kettei-chub, kettei-core. Docs price kettei fat at $0.06 per million input tokens. Opened October 17, JevBench API board v1.7.46 ranks kettei-fat first of 39 at composite 77.8. GET /v1/models with no key returned HTTP 401. This desk did not send a billed POST.
respanai/span-02
Gemma 4 26B-A4B fine-tune, about 4B active, NVFP4/FP8 16.4 GB. Card prints Decision Index 0.3 text index 63.62. Local POST /v1/systemone on the included engine. Hosted POST https://api.respan.ai/api/v1/decisions, input_cost 0.01 on the public models list opened October 16, 2026. This desk did not run the weights.
OpenRouter Armor
Beta plugin posted October 15, 2026. Scores tool calls with a decision model before they run. The 492-call chart prints Jev 1.13 at 90.7%, 446 of 492, $0.0021 a call. This desk did not send an Armor request.
SoundBlaster/Jev4Mellea
Apache-2.0 adapter. Tag v0.7.0 at 3b1bdf68a1e0, 13:52 UTC October 15, 2026. Experimental PNG/JPEG inputs for Noul, Choice, and Score. The October 15 visual comparison records twelve scenes on jev-1.14.0, coverage 5 of 12 at 0.10 / 0.90. Bounding boxes sit outside this adapter's image interface. This desk did not install the package.
brainstormity/JEVsnap
MIT screenshot redactor, v0.2, commit b249546b2538 at 05:54 UTC October 15, 2026. Default AI path sends OCR text with local secrets replaced by [HIDDEN]. Box visual check is off by default. This desk did not install the app.
truefoundry/trueforge
MIT agent harness. npm @truefoundry/trueforge 0.178.0, opened October 14, 2026, requires Node 22.14 or newer. Akshay Pachaar's weekly client follow-up agent scored 37 of 40 with Jev and 37 of 40 without it. He wrote that 40 is too small for the confidence cut. This desk did not run npx @truefoundry/trueforge.
sitammeur/trueforge-jev-backlog
Synthetic GitHub issues for that article. README prints 300, 200 open and 100 closed. Opened October 14, 2026, a GitHub issue search returned 306. The license field was empty.
Celeris-1 Decision
Hosted decision model posted October 8, 2026. POST https://inference.celeris.ai/celeris-1-decision/v1/systemone, model celeris-1-decision, $0.04 per million input 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. GET /v1/models with no key returned HTTP 401. This desk did not send a billed POST.
Liquid d1-ultra
Hosted decision model posted October 15, 2026. POST https://api.liquid.ai/decisions/v1/systemone, model d1-ultra, $0.042 per million input tokens, output free. Opened October 16, Artificial Analysis Decision Index prints 65 against Jev 1.14 at 57. This desk did not call the API.
SubQ Snap
Hosted decision model posted October 12, 2026. Docs name POST https://api.subq.ai/v1/systemone, model subq-snap-preview, input $0.05 per million tokens, output free, window 384K. The blog prints Decision Index 0.2.1 at 61.7 against a Jev cell of 57.9. On October 13, 2026, at about 03:00 UTC, two POSTs from this desk returned HTTP 200.
chinhnc/Quyet-1.0-Large
Apache-2.0, Gemma 4 31B, 31.3B on the card. The homepage prints frozen JevBench v1.6.0 at 71.39 and Capability 81.7. Four smaller checkpoints are in the same family. We did not run the weights.
PixilabAI/Blink-v0.3
NVFP4 checkpoint on Gemma 4 26B-A4B. The card's license field is Apache-2.0. Decision Index 0.2.1 on the lab's run is 57.48, against the board's Jev cell of 57.91. We did not run the weights.
vllm-sr/d3-ultra
Apache-2.0, from Qwen3.8-27B, 26.09B on the card. Opened October 26, Decision Index 0.4.1 Full score is 70.18, second of 229. The card's private SR-Text mean is 87.2. We did not run the weights.
vllm-sr/d3-turbo
Apache-2.0, from Qwen3.8-27B, 26.09B on the card. Opened October 26, 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. We did not run the weights.
vllm-sr/d3
Apache-2.0, from Qwen3.8-27B, 26.09B on the card. Opened October 26, 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 now aliases this id to d3-turbo. We did not run the weights.
xmanii/Mizan-4B
Apache-2.0 decision head on Qwen3.5-4B. Mean 78.7 on 3,000 real Persian records, against the card's Jev cell of 83.7. The Q4 GGUF is 2.5 GB at 78.3. Text only. We did not run the weights.
doofz/THX-01
Apache 2.0, 322M, on mmBERT-base. Support-ticket average 98.4 against Jev 1.13 at 97.4 on 2,843 tickets. The docs name POST https://api.hal-x.ai/v1/systemone. We did not call the API.
h2oai/h2o-lightning-4b
Apache 2.0 checkpoint from Qwen3.5-4B, named on H2O's October 7, 2026 blog. That post prints JevBench composite 72.5 against Jev 1.13.0 at 71.5. The blog says the board measured v1.1, and that revision v1.2.1 adds images. We did not open the card or run the weights.
nace-ai/drex-v1.5
Open weights posted October 9, 2026. About 9B, Nace.AI Open RAIL-M. Decision Index 0.3.1 public index 58.08 against Jev 1.13.0 at 57.96. Commercial use above $1,000,000 of revenue or funding needs a separate license. This is not the Drex DLM card. We did not run the weights.
stas4000/openbles-firewall
MIT adapter, 31.8 MB, on fastino/GLiNER2.5-Decide. The README scores 119 of 120 agent attacks caught and 0 of 120 benign look-alikes flagged. We did not run the weights.
levantolabs/sagecraft
Harness that shows Levanto's Sage a screenshot and a short list of moves. The October 9 post logs 4,573 decisions. Sage itself is a hosted API, ranked first at 74.0 on the October 11 JevBench API board. We did not open this repository.
Cloudflare/clef-omni
Apache 2.0 weights from Qwen3-Omni-30B-A3B-Instruct, posted October 9, 2026. The blog's BANKING77 cell is macro-F1 94.8. Hosted price on that post is $0.15 per million input tokens. We did not run the weights or call the API.
olaverse/PurpleMIST-Flash-1.0
Apache 2.0 checkpoint from Qwen3.5-9B-Base, text tower only, about 7.9B parameters. The card scores 0.668 accuracy on 15,050 translated African-language decisions. serve.py answers POST /v1/systemone. We did not run it.
NandhaKishorM/vegaml
Apache 2.0 package. Default checkpoint is 0.8B on frozen Qwen3.5-0.8B, with a 4B sibling. The README scores 75.4% on 800 phishing emails against Jev 1.13.0 at 61.9%. This is not Decision 2.0's Vega. We did not run the weights.
TypeLLM/TypeLLM
Apache-2.0 library that constrains an autoregressive model with a JSON Schema. PyPI typellm 0.6.8 was uploaded October 9, 2026. The September 23 file scores RadixArk/Qwen3.8-27B-NVFP4-BF16-LMHead at 228 of 231 public JevBench tasks with thinking, and 195 of 231 without it. The hosted name typellm-latest is not in that file. We did not serve the checkpoint or call the API.
Microsoft-Decision-1
Hosted decision model posted October 9, 2026. The catalog, opened October 10, says the weights are not distributed. Post-trained from Qwen3.5-9B. Context 32,768. Lifecycle label Generally available. Command Line prices input at $0.042 per million tokens. Foundry path /providers/microsoft/v1/systemone. We did not call it.
bandr-ai/bandits
Step-Jev, posted October 8, 2026. The recipe trains Qwen3.5-4B-Base LoRAs that score a search step and a final answer. On one BrowseComp-Plus seed the step reward scores 19.6% against 5.4% for the final answer alone. Adapters from that study are written to a Modal volume, not a Hugging Face repo we found. The GitHub license field was empty on October 9. We did not train it.
nace-ai/drex-dlm
Open weights posted October 6, 2026. The post calls them Drex 1.1. The card title is Drex DLM. Backbone NVIDIA Efficient-DLM-8B. POST /v1/systemone. Weights CC BY-NC 4.0. Nace code MIT. Decision Index 0.2 cell 52.31 against the kit's Jev cell of 51.67. nace-ai/drex-agent-skill points at this server. It is not a second model.
musubilabs/policylm-1.7b
Apache-2.0. Fine-tune of BidirLM-1.7B-Embedding. A score from 0 to 1 for each listed category, up to 16 in one pass. Custom-policy accuracy 0.842 at a 0.5 cutoff. The table has no Jev column. The card does not serve Choice, Noul, or /v1/systemone.
vllm-project/semantic-router
The Decision 2.0 cards name this repository. Scores are on the Hugging Face cards, opened October 5. Vega prints Decision Index 56.5 and does not print a Jev row. Apache-2.0. Six checkpoints, Kai through Vega. Decision 3.0, posted October 11, is six more Apache-2.0 cards from the same project. d3 prints 64.7 on its Decision Index 0.3.1 table.
vllm-sr/Vela-2.0
Posted October 7, 2026. Four Apache-2.0 checkpoints, 0.3B through 9B. The 9B card prints Decision Index 0.2.1 at 41.63 with Noul calibration on, and 41.09 with it off. The evaluation file prints Lux at 46.23 and does not print a Jev row. The 0.3B tokenizer carries the Gemma Terms of Use. The blog's example server is POST /v1/systemone. We did not run it.
unslothai/unsloth
October 7, 2026 documentation for a Clef head on a public LLM. Qwen3.5-0.8B moves from 36% to 73% on typed-decisions. The same page prints Llama 3.2 3B at 79% and Gemma 4 E4B at 77%. We did not run the notebook.
togethercomputer/tev1
Training recipe linked from the Tev1 card. Ollama's page says the dataset builders and training scripts are MIT. The Hugging Face card, opened October 7, 2026, says the fine-tune license is still being finalized. Together's page that day lists $0.04 per million input tokens. The chat call returns one letter.
dreamerron/decision-bench
README says MIT. 450 hand-labelled harness decisions. fsdecide, a 149M ModernBERT, scores 91.3% on the 80-item review suite against jev-latest at 56.3%. On the authors' 1,248-request general set, jev-latest is 90.5% and fsdecide is 60.5%.
autotrust/JEV-27B-VL
Apache-2.0 adapter on Qwen3.8-27B. POST /v1/decide. October 3 section: 15 of 20 pick-and-place scenes, 95% of 60 browser tasks when element text is included, 10% with numbered boxes only. Six-benchmark mean 84.07 against hosted Jev at 83.85. That JevBench column is not the official board.
StartLuxLabs/StartLux-Decision
Code Apache-2.0. Weights CC BY-NC 4.0. README opened October 20: v1.1 27B Instant 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.
rkinas/basal
Apache-2.0 engine. GitHub main on October 4 still describes basal-1.0. Opened October 8, 2026, that README is titled basal 1.5. The basal-1.5 scores are on basal.si5.pl. Werdykt v1 puts the 4.5B at 72.12% and Jev 1.13.0 at 81.64%. The site schedules the 1.5 weights for October 5. Tag v1.5.0 returned 404 that day.
eulogik/nirnay
Apache-2.0. NIRNAY 450M, Laya 421M plus about 30M. Weights at eulogik/nirnay-450m. phase_b JSON is 0.8792 on 3,080 Banking77 rows. The card says the Jev cell was not measured in the repo. POST /v1/systemone. Context 512 tokens.
SamratDuttaOfficial/WaterSheep
Apache-2.0. ModernBERT-base fine-tune. POST /v1/systemone on port 8766, plus a multi-label type. README prints 77.8% in distribution, ECE 0.026, and 61.2% on held-out datasets, ECE 0.043. No Jev column. Card version 0.1.0.
BudEcosystem/Bud-Decision-Engine
Desktop app for eleven open decision models. Local POST /v1/systemone at 127.0.0.1:8420. GitHub lists no license. The GB10 training table prints Jev-Omni at 62 percent to 77 percent on business workflows and does not print an example count.
strands-labs/strands-decider
Apache 2.0. Strands Decider 2B, 1.9 billion parameters, checkpoint StrandsAgents/strands-decider-2B-hobson-v19. v19 scores 167 of 231 on the public tasks, median 115 ms on an RTX 3090. POST /v1/systemone. The rank in the file is the September 25 v1.4.2 board.
blockbrain-ai/cygnet-recipe
Cygnet, frozen Gemma-4-12B-it. JevBench v1.5.0 official rank 1 at 73.7. The credit line says the shim is MIT and the weights are Apache-2.0 under Google's Gemma policy. Capability 79.0, median 0.23 s.
EldanRing/Winnow-12B
Winnow-12B Q8. JevBench v1.5.0 official rank 2 at 73.2. The page calls this a statistical tie with Cygnet. The credit line says Apache-2.0, including the Gemma 4 terms. Capability 79.3, median 0.34 s.
ollaya-dev/ollaya
Apache-2.0. Local server on port 11435, POST /v1/systemone. The README says it is independent of Ollama and of TypeSafe. winnow:e4b is 0.722 against TypeSafe Jev at 0.738. The item count for that pair is not defined on the page we read.
Maincode/matilda-jev
Card license line Apache 2.0, 26.1B parameters. Self-run of Decision Index 0.2.1 prints 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. The page we opened is the Hugging Face repo matilda-jev-v1.
PostHog/jeeves
MIT. Qwen3.5-9B LoRA plus a pointer head. Public JevBench hard, sealed tier left out, is 0.865 against a Kev-published Jev cell of 0.730. Transfer overall is 0.746 against Jev 0.800.
mohit67890/imajev
Apache-2.0. Qwen3.5-4B LoRA, frozen vision tower, up to two photos. Changelog v1.4.2.2 ranks Imajev-4B first at 67.37. On the v1.5 addendum it would place sixth at 70.4, outside the ranked order. 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.
crh225/plumb
Apache-2.0. Starts from JevK5 v0.2. Public-hard runner on one RTX 4080 Super: 89 of 111. Official composite 65.84. Hard ECE 0.094, worse than v4's 0.051. The v1.5.4 page numbers 71.6 as official rank 5, still tagged roster addendum.
skundu42/kev
400M encoder, Choice, Noul, and Score. No LICENSE file in the root listing. Not jaredpalmer/kev. 79.96% against Jev 73.54% on 17,476 of its own rows. PhishNChips 50.25% against a published 62.6%.
feder-cr/jev
jevos-1b, Q4_K_M, llama.cpp, POST /v1/systemone. Noul only. README alt text: 0.815 against Jev 0.927 on 2,000 yes/no questions. pyproject says MIT. No LICENSE file in the root listing.
Contrastive-LM/CLM
Apache 2.0. Frozen Qwen3-8B, two 20M heads, POST /v1/systemone. DeepSWE verifier holdout 31/38. Terminal-Bench 2.1 printed as 87.6% of 30 tasks.
caiovicentino/eikos
MIT code. Eikos-4B and Eikos-27B, letter logits, POST /v1/systemone. Author hard-tier row 82.9 against Jev 73.0. Error at 0.90 confidence is 2.4% on 7,140 items.
caiovicentino/eikos-arena
MIT. Paper trading on 14 Hyperliquid perpetuals, $10,000 each, 28 questions every 5 minutes. File opened at 09:20 UTC on September 26, cycle 626 of 864: Jev equity $10,051.22, Eikos $9,653.19. The September 25 file, cycle 302, was $10,125.97 and $9,922.72. The 72-hour run ends September 27.
TheoLeeCJ/SemIf
Formerly OpenJev. Reads option logits from Qwen3.5-4B. JevBench 74.6 versus Jev 75.3. Independent; not affiliated with TypeSafe.
githubnext/localjev
Bun server that speaks POST /v1/systemone and prompts a local model on oMLX for JSON probabilities. GitHub Next. Not a logit read.
logan-markewich/jeff
Self-hosted System One API on GLiFormer. 1,600 public items against live Jev.
bespokelabsai/nimble
9B Qwen LoRA trained to imitate Jev-style answers. 90.12% on 324 synthetic labels, 74.8% macro on 13 public subsets.
Heman10x-NGU/Verdict-open-jev
151M encoder that runs in a browser tab. Cited next to Nimble, not as its own story.
NandhaKishorM/laya
Apache 2.0 ModernBERT-sized encoder, three checkpoints, a Router. JevBench 70.1. Author Banking77 0.425 versus a published Jev 0.870.
mizorewww/laya-mlx
Apache 2.0 MLX port of the Laya checkpoints. M3 Max one-question median 13.42 ms (421M) and 7.39 ms (322M). Selected answers matched upstream on 63 of 63 validation questions. Not a new JevBench row.
jaredpalmer/kev
Apache 2.0 LoRA plus pointer head on Qwen. Serves POST /v1/systemone. Author new-source: Kev-8B 79.6% versus Jev 85.7%. JevBench kev 0.6B 66.7. September 25 post: Kev-4B on OpenRouter at $0.042 per million input, $0 output, 8K context. That model URL returned Not Found on September 26.
oaleynik/kev-jev-eval
September 26 paired run. Pinned Kev-4B 139fdd94 against a fresh jev-1.13.0. Full JSON function selection 1,216/1,253 versus 1,235/1,253. Prose parameters 1,059/1,253 versus 1,236/1,253. Kev is loopback HTTP. Jev is hosted HTTPS.
Davipar/djev-dev
Apache 2.0 local POST /v1/request on DiffusionGemma plus patched vLLM. Hosted API is Maisa's djev.dev. JevBench 74.3.
mmastrac/djev
Apache 2.0 example server from vLLM pull 57250. Seeds a canvas and reads noul, choice, and score after one denoise step. POST /v1/systemone. A hand chart adds to 198 of 201 against Jev at 191 of 201.
NakliTechie/sieve
MIT. Zero-shot arm is DiffusionGemma-Jev. Reversing the options flipped 24% of answers. Three shuffled reads halved that and added 17 to 26 points on sets with 60 to 151 options.
Hangzhi/diffusion-jev-sglang
September 21 replay on one A100. Public JevBench tasks 196/231 against the published Jev artifact of 200/231. Text ECE 0.1222. The read is the last active denoising step, often two, beside the one-step vLLM server.
Zefan-Cai/Open-Jev
MIT source. Qwen3.5 2B and 9B, plus a Qwen3.8-27B v1.1 package, LoRA and a decision head. Apache 2.0 adapters. 231 public JevBench tasks: 9B 179/231, 27B 197/231, Jev 200/231. Different project from SemIf. An October 2 note, frontier-controls-v4.md, records one 64-step pilot on the released 2B, 84/128 to 99/128 on synthetic rows.
featherless-ai/simple-jev
Apache 2.0 classifier server. Reads next-token logits from a Hugging Face chat model. JevBench Qwen3.8-27B 67.3.
NakliTechie/llamacpp-jev
MIT. POST /v1/systemone in front of unmodified llama-server, text and images. Raw label softmax. Fresh 448 image, four questions, 526 ms median on an idle M4 Pro.
kishida/llama.cpp
MIT fork, branch jev. The branch builds POST /v1/systemone and GET /jev into llama-server. Opened again October 6, the page still says the endpoint is not upstream. The newest commit on docs/jev.md is October 1. The table is 1,191 questions. Qwen3.8 Flash-Next is 0.944, marked Mac, 292 ms.
ggml-org/llama.cpp pull 29818
Merged October 2. llama-server answers POST /v1/systemone for Julia-1, Laya, Kev-4B, lev, and OpenJev. OpenJev is the image model in that note, CC BY-NC 4.0. The other four are Apache 2.0. Median one-question times on one RTX PRO 6000 are 3 ms, 5 ms, 12 ms, 36 ms, and 43 ms.
ggml-org/llama.cpp v0.6.0
Published October 5, 2026, nightly b11429. POST /v1/systemone adds Nimble and Clef. Nimble is Bespoke-Nimble-9B-v3, Decision Index 0.2.1 at 56.88, CC BY-NC 4.0. Clef can take an image. The October 6 note leaves Clef's speed cell blank.
54yyyu/rev
MIT. Frozen open model, option logits, POST /v1/systemone. Qwen3.8-27B public hard tier 0.784. Laptop Qwen3.5-2B hard tier 0.550. The README's Jev hard cell is 0.730.
abhishek085/open-spark-jev
Apache-2.0. spark-s1-4b-v6 is a Qwen3.5-4B LoRA. Public-proxy Intelligence 83.1. Hard tier on that page is 0.595 of 111. Not the official composite.
mrmps/tinker-binary
Apache-2.0 training code for hotdog-27B. The adapter is on Hugging Face at opensporks/hotdog-27B. Binary logit read. Holdout 1,603/1,631.
taeold/djev-run
Cloud Run recipe for mmastrac/djev-spark on an RTX PRO 6000. POST /v1/systemone. Samples=1 latency 35 to 60 ms. No license listed. No accuracy table.
Rizzo-AI-Academy/rizzo-flow
Apache-2.0. Spark-X2.5-4B on llama.cpp, POST /v1/systemone. Q8_0 p50 49 ms on an RTX 5060 Ti. authored144 0.812 against SemIf's published 0.819. 26 options.
trotsky1997/jevre
MIT. 0.85B checkpoint, TypeSafe Python SDK on localhost. Test accuracy 0.9323. Held-out sources 0.8366, ECE 0.0944. Served p50 77 ms.
Argos1111/jev_local
Local /v1/systemone for LFM, Sarashina2.2 Vision 3B, and a ModernBERT-Ja cross-encoder. JGLUE test: Sarashina Q4 34% / 86% against LFM 17% / 69%. No LICENSE file.
abxda/ambar-reflex-sistema-1
MIT. Qwen3.5-4B plays an original run-and-gun from six typed questions. Realtime progress 24.9% at 408 ms. Offline 9B accuracy 82.7% on 554 decisions.
Loops
rhighs/jev-code
TypeScript SDK. Jev picks grammar productions and local code builds the AST. Ladder on commit 517d0f0 fails three tasks. Hello world is 3 turns and 16 requests. No LICENSE file in the root listing.
browser-use/jev-ultrafast
Browser Use's indexed action-space agent. Zurich to London Google Flights at 7.1 seconds in the repo notes.
jkudish/jev-browser
Stagehand-related browser loop. Cited in the computer-use story.
jcpsimmons/jev-macos-loop
Native Mac window controls. Finder batch of nine files in 7.39 seconds through Vercel.
jaewgwon/jevis
Flutter integration_test package. Rank registered widget actions. Strings come from the test author.
maxlibin/jev-rubiks
MIT browser cube. Kociemba solves in code. Jev decides whether the coach speaks. A removed test of 18 turns, forty times, left the distance at 22.
idovmamane/dejevu
MIT browser loop. One JSON action to a chat model. Groq flights median 5.63 s against jev-ultrafast's published 7.09 s. The Jev path in the repo is untested.
Jinzhengxu/poker-table
GPL-3.0 Hold'em table. Jev reads the opponent in one request. Code picks fold, call, or raise. Spot-bank median about one second on 46 spots.
standardagents/jevpilot
Three.js driving sim. Jev picks among sampled steering paths.
elalber2000/jev-rlm
MIT fork of alexzhang13/rlm-minimal. One 100,000-line needle, seed 42. 82 Jev calls and 1,081 seconds, or 78 recursive calls and 1,480 seconds. Both answers are 2867825.
awlevin/typesafe-computer-use
MIT macOS loop. OCR, then Jev chooses the action. One shared screenshot: $0.0002 against Opus 5 at $0.032. Dates are parsed in code. No task success rate.
Routers and filters
ohernandezdev/jevmod
MIT, 0.2.1. One Noul per category, including nsfw at 0.80. Sexual AUROC 0.984 on OpenAI's moderation eval, F1 0.866 at that line, remeasured September 23. $0.044 per 1,000 messages is the whole request.
mohamadkhoshnava/ZeroNSFWBot
Rust 0.1.0. The README says MIT. GitHub's license field is empty. Pictures stay on the host: Marqo's ViT-tiny, then giacomoarienti's five-class ViT. Jev reads text at jev-latest when a key is set. No printed accuracy on adult images.
stanfordnlp/dspy
MIT. Release 3.4.0, September 25, adds dspy[typesafe] and experimental Noul, Choice, and Score types. Predict sends decision signatures to TypeSafe. The notes print no accuracy table.
cmpnd-ai/dspy-system-one-agent-patterns
Drew Breunig, October 1. Pins dspy[typesafe]==3.4.0 and TypeSafe jev-1.13.0. The shell gate refuses below 0.2 and runs at 0.6 or above. GitHub lists no license. The README does not print a fitted threshold or a v5 score.
VectifyAI/jev-doc-search
Apache 2.0. jev-1.13.0 walks a PageIndex tree when one Choice would pass 32k tokens or 255 options. The printed table is two filings. Both are marked correct, NVIDIA page 51 and Citigroup page 134. We did not run the scripts.
yctimlin/JevScout
MIT. Claude Code hook. Condenses MCP results of at least 100 KB. The README's check is 38% less agent time and 26% less agent cost. TypeSafe's token bill is separate. Not an official TypeSafe package.
championswimmer/pi-skill-picker-jev
MIT, package 0.1.0. Pi extension. Asks a decision model which skill names to load. The 20-skill README demo cuts about 1,000 tokens. If the service is down, it does not load the full catalog.
dzhng/jevgrep
MIT. Jev picks files and excerpts. Sol still writes the patch. Package 0.4.3, ten tuned Python tasks: 8 of 10 official solves on both arms. Combined cost including a Jev list-price estimate is 25.8% under a saved baseline.
kyu1204/jgrep
MIT. Command jgrep, npm name jevgrep. A different repository from dzhng/jevgrep. The September 28 bench is mean recall 0.926 on 60 commits, selection 0.119, total $0.1112. The model field sent is jev-latest.
RiskAverseTech/toolgate
MIT. Static rules first, then Jev scores the call. Set 6 on 0.15.0 is 19 of 20. Not diorrego/toolgate-experiment, and not a sandbox. The 1,369-decision window is version 0.9.1.
Oranquelui/astra-jev-harness
MIT. Asks Jev which source bytes a Codex prompt still needs. September 24 fixture: Astra input 19,198 to 14,011, checks 6/6. September 22 files 28 to 4, Astra input down 1.2%.
yusukebe/hono-jev-router
Hono Worker that asks Jev whether a request looks like spam, an agent, or an angry customer.
tamaratran/fast-jev-compaction
Score tool-call blocks and drop them instead of summarizing.
GhalebDweikat/winnow
Similar filter ahead of Claude Code.
SamSaffron/term-llm#1150
GPT Live control plane. Shadow mode on. Fail open to steer.
m-mizutani/semgate
Go net/http middleware. Asks Jev about each request. Default sends every header, including Authorization and Cookie.
KranzL/Jevflake
dbt package plus Terraform. Snowflake SQL functions for noul, choice, and score. Calls TypeSafe's API; row content leaves the warehouse.
BerriAI/litellm
Proxy pass-through at /typesafe/v1/systemone in v1.103.0-rc, plus a typesafe compaction guardrail at relevance_threshold 0.2.
d-date/swift-jev
MIT Swift package and CLI. Choice, Noul, and Score in one call. Sample policy: escalate under 0.60, act at 0.85. No scored set.
alexforman1/postgres-search
MIT. Postgres search with an optional Noul per top-10 hit. SQL first-place rate 82% on 50 queries. The Jev pass is unmeasured. Default threshold 0.30.
mustafakendiguzel/aso-workbench
Mac app. MIT with the Commons Clause. Jev features feed a local popularity model. Spearman 0.74 on 578 held-out terms.
Nasrallah-AL/jev-cli
npm jevctl, command jev. verify, screen, classify, compact, batch. TypeSafe, OpenRouter, or Cloudflare. compact keep-threshold 0.5.
jerryjliu/docjev
Apache 2.0 classify and split for PDF, DOCX, PPTX. LiteParse locally, hosted Jev for the decision. 40/40 classify, 7/8 split versus Luna.
khmuhtadin/n8n-nodes-jev-classification
MIT n8n community node. Classify, Score, Check, Ask Questions. One output per category plus Needs Review. Self-hosted; n8n declined to verify it.
realZachi/pg-jev
PostgreSQL License. Extension name jev on PGXN. WHERE jev(table, condition). README: 2,000 rows in about 3.5 s and about $0.012, then about 50 ms from cache. Batch 20. Needs plpython3u and a superuser.
mejiasd3v/pg-jev
MIT PostgreSQL extension, a different project from realZachi/pg-jev. prompt_jev returns jsonb from TypeSafe, Vercel, or Cloudflare. No scored set in the README. One HTTP call per row.
backant-io/jevelry
MIT npm runtime. A JEVEL.md holds the questions and the act / mark / fall_back thresholds. Last live note: three calls in 2.49 seconds on jev-1.13.0.
prasanthj/duckdb-jev
Apache 2.0 native C++ DuckDB extension. Batches and streams Jev from SQL. Live Choice: 2,049 rows in 0.887 s. Row text leaves DuckDB.
colliber/duckdb-jev
Separate MIT native DuckDB extension. Bind-time ENUM and STRUCT result types. One request per row. Related work on the Prasanth story.
assadiandre/jev-search
macOS file search, no saved index, no LICENSE file in the root. JEV reranks the top 128. nDCG@10: NFCorpus 0.322 to 0.377, SciFact 0.687 to 0.745. 623 queries, 66.3¢.
kylemclaren/jevsearch
MIT shadcn registry block. Keyword hits first, then Jev re-ranks the top 20. Author bench 83% Hit@1 on TypeSafe docs.
aifabrice/jev-rag
MIT local search. Default path is SQLite BM25, then Jev, then MiniMax. NFCorpus best headline row is nDCG@10 0.451. The four-question Passage Gate scored below ordinary reranking. The README says the project is not affiliated with TypeSafe.
ajanm007/jevrag
MIT. Five RAG decisions behind one interface, Jev first. HotpotQA n=700 matches a three-round baseline at 38.5% fewer retrievals. ECE on that sufficiency question is 0.3322. The HotpotQA split is not in the public repo.
softwaredoug/search-experiments
Bag-of-decisions runner. WANDS CSV mean NDCG@10 0.574262 for one Noul and 0.609814 for GPT-5 questions, both medians 0.5609447680982702, commit 8734fde. Score is retrieval plus 10 times the summed Nouls, top 100. The GitHub license field was empty on October 9.
vectorize-io/hindsight
0.10.1 adds a typesafe reranker. Listwise Choice. LoCoMo, 30 candidates, recall@1 0.950 against local MiniLM at 0.800. Pruning is off by default.
keltokhy/jsort
MIT pairwise ranker, command jsort. Bradley-Terry scale from Jev Noul comparisons. CommonLit r 0.824. Fed script is 95 statements.
Engid/question-kit
MIT, package question-kit 0.2.0, author Nick Gideo. Commit 25a3b71 on order-v2-and-core-updates adds rankByScore and rankPairwise, listed as Unreleased. rankPairwise is a two-option Choice per pair plus Hunter MM Bradley-Terry, prior 0.1.
vercel-labs/jev-ai-sdk-form-router
MIT Next.js template. Accepts Jev at confidence 0.95. GPT-5.6 Luna Fast decides the rest from the submission alone.
syumai/jevyoumean
MIT Go wrapper, command jym. Unknown subcommands become one Jev Choice against help text. Valid commands do not call the API.
vinilana/jev-gateway
MIT local proxy, npm jev-gateway. Jev picks the tool for Codex, Claude Code, OpenCode, or Gemini CLI. Direct mode can skip the language model when every argument is closed.
spring-ai-community/spring-ai-typesafe
Apache 2.0 Java SDK and Spring AI advisors. Judge, guardrail, document filter, tool index. Blog pins 0.1.0. README quick start still shows 0.1.0-SNAPSHOT. Opened October 8, 2026, the quick start pins 0.4.0, and Maven Central lists that release. GitHub release 0.3.0, September 30, points the same client at Laya. The Ollama notes are on the snapshot docs. A Maven search on October 1 still returned no documents for spring-ai-starter-typesafe.
everruns/everruns#3924
Merged September 30. An opt-in OpenAI decision driver over POST /v1/decisions. That pull request's live smoke returned HTTP 403, Decision API is not enabled for this user, and it keeps guessed types behind a preview flag. On October 6 this desk's own POST returned HTTP 200. We did not rerun the pull request.
use-agent-os/agent-os#3316
Opt-in Jev strategy for Pilot Router. Not the default. Chart on 121 labeled turns: Jev accuracy 0.901, local MiniLM 0.355. high_risk at 0.7 floors the turn at c3.
xinyao27/jevonian
AGPL-3.0 local proxy on port 8787. jevonian/auto asks Jev once for the route and the thinking depth. A pinned model id skips the call. No scored table in the README.
Tech-Byte-Frontier/jevgate
MIT or Apache-2.0. Rust review gate, crates.io v0.5.1. One function per question. Findings compose at 0.80. The README says the probabilities are judgments, not a measured accuracy.
Measurements and catalogs
AppliedMachineLearning-Lab/jev-benchmarking
Named in the abstract of arXiv:2609.37647. 37 datasets, 346,009 requests, under $10. We read the abstract and did not open this README.
arXiv:2609.37647
Evaluating and Benchmarking the System One Model Jev. Submitted September 29. Tobias Deußer, Lorenz Sparrenberg, and Rafet Sifa. Jev beats Qwen3.8-27B on 27 of 37 and Gemma-4-E4B on all 37. Responses are on Zenodo 10.5281/zenodo.23039006.
cua-speedrun/cua-speedrun
The jev-agent tree linked from the October 2 chart. The printed callout is 9.8% on 50 OSWorld tasks at 125 seconds each. We did not run it.
StreamDecisionBench
Jacob Lin, October 2. The thread points here for the leaderboard and the code. The HTML we opened did not render the table. The chart image prints Kev-27B at 62.15 and the Jev API at 59.63. arXiv:2609.38612 is a different scoring.
arXiv:2609.38612
Jhen-Ke Lin and Chung Chun Wang, NYCU. Submitted September 29, version 2 dated October 1. Do not mix that table with the October 2 chart.
youdotcom-oss/risk-analysis-server
MIT. Edward Irby, October 1. Bun. Three escalated sweeps total $0.43, or $0.14 each. Jev is the typed judge. Qwen3.8 27B proposes queries and writes the briefing.
ZhangYiqun018/jev-dimabsa
Frozen jev-1.13.0 and 488 CPU-fitted coefficients for SemEval-2026 Task III Track A. Table 1 prints Task 1 RMSE 1.0645, Task 2 continuous F1 52.09, Task 3 44.06. We did not open a LICENSE file in the repository.
arXiv:2610.02048
HydroJEV, submitted October 1. jev-1.13.0 on simulated water-network alarms. Mean in-distribution macro-F1 0.63. The agreement gate settles 69 of 200 and 76 of 200 fresh windows. CC BY 4.0. We did not find a code repository in the sections we read.
AI-secure/DecodingTrust-Agent
The DTap platform named on Zhaorun Chen's chart. The public leaderboard listed 0 agents when we opened it on September 29. The Jev percentages are the post's image. The May platform paper does not contain those cells.
sumleo/RLCDAlignBench
44 benchmarks, 7,193 instances, jev-1.13.0. Generic Noul median AUROC 0.886. Median ECE 0.168. Code MIT. Data CC BY-NC 4.0. Hugging Face copy is gated.
Robertzu43/system-one-security-triage
Recorded 100-case security demo. Jev, Terra, Opus, five passes. Dashboard serves 2026-09-19-public-v5.
fstandhartinger/jevbench
JevBench harness and frozen tasks. v1.5.0 ranks Cygnet 73.7, Winnow-12B Q8 73.2, and Jev 72.1 on 1,624 decisions. The older v1.4.2.2 board keeps Jev at 63.29 and Imajev-4B at 67.37. The v1.2.8 geometric mean was 75.4. Not affiliated with TypeSafe. The page opened September 30 is v1.5.4 and numbers 106 systems. Those three scores stay on top. 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.
Mapika/decider
decider-4b v2. The v1.4.2 row is 64.13, Intelligence 49.4 against Jev 53.1. On v1.5 the same row is official rank 5 at 71.3. Cost on the row is an estimate. The board's note says Apache-2.0. On v1.5.4 that row is rank 7 at the same 71.3.
shapor/jev-sentinel
Named on the QueryStory demo page. We did not open the repository. TechCrunch prints $2.94 against $372 for one monitoring job. The demo page prints about $55 per million actions and does not print those two figures.
LiquidAI/d1-3B
Open d1, posted October 7, 2026. d1-3B is 3.12B, from LFM2.5-VL-3B. The card prints Decision Index 0.2.1 at 48.57, self-scored, not a board submission. Winnow-12B on that table is 50.02. d1-omni-600M, a second card, prints 15.95. LICENSE is LFM Open License v1.0, commercial use under $10,000,000 annual revenue. The GGUF card prints llama-server -hf LiquidAI/d1-3B-GGUF:Q8_0 and POST /v1/systemone. We did not run it.
apolinario/decision-index
Reproduction kit for Decision Index 0.2. Forty benchmarks. The test locks Jev at 51.67. Code is MIT. The suite text stays under the upstream licenses. Maincode's September 30 card prints a self-run of the 0.2.1 kit at 59.26 for Matilda. That 59.26 is not in the September 28 board file. Liquid's September 29 chart is a separate reproduction at 58.9 against 57.9. Nace's chart prints Drex 1.5 at 199 of 231 against Jev at 201 of 231. Fastino's October 1 blog prints GLiDE at 64.81 on another self-run. AutoTrust's October 3 card prints GEV-26B-Decide at 62.48. Neither name was in the September 28 file when we last opened it.
SamanPandey-in/jevrail
MIT hook in front of a shell command. Sample deny band 0.80. Printed eval is the offline corpus: 22/22 catastrophic, 15/26 false asks.
confident-ai/deepeval
Apache-2.0 LLM evaluation framework. JevEval sends Noul, Score, and Choice questions to Jev in one request and returns a weighted mean. PyPI deepeval 4.2.8 uploaded October 2, 2026. This desk applied the published equation to two hosted jev-1.13.0 calls and did not install the package.
danielgshea/jev-as-a-judge
LangChain weather-agent judge eval. Frozen traces, oracle labels, 100-repeat variance.
arXiv:2609.26550
Yubo Li and coauthors, v3 revised September 29. JEV is 1.7 points behind GPT-6 on 4,850 public judgments. A frozen cascade is 0.9 points above GPT-6 on 1,610 pairs. A different project from the LangChain weather-agent repo above. We did not open a code repository for this paper.
arXiv:2609.31142
JevAdvBench, submitted September 25. The abstract says one appended opinion flips 12.1% of jev-1.13.0 decisions. We read the abstract, not the 19 tables. A different paper from the 510-case InjecAgent attack.
onlyoneaman/jev-eval
Public half of Kumar's 16,000-call write-up: cases, per-item answers, scoring.
SREGym/SREGym
SREGym-Lite harness. 20/50 without Jev, 24/50 with it, on 10 problems times five attempts.
elvisun/newsjack
News Desk Dealer demo. 384 headlines versus Opus 5.
jarrodwatts/jev-trader
Monad/Kuru buy-or-sell loop. Mock model unless you set a key.
hellogumbo/awesome-jev
Community directory. Week-one story lists a slice of it. Counts keep moving.
vinilana/jev-gateway-bench
Chess-engine sessions, routing on and off. 120 runs in the gateway README. Astra and Sol bug-fix output -57%. Opus 5 feature task +83% time.
hw725/dansa-research
Jev as a fourth judge on 28,074 translated sentences. Gaps +38.3, +11.3, +36.8 points. About $0.54. No gold label. No LICENSE file in the repo root.
7shi/multilingual-reader
CC0. Jev 1.13.0 scores 16 translation models on 67 languages, one run each. 1,072 evaluations, 7,536,547 input tokens, $0.3165, 5m22.6s.
TheWayWithin/jev-bench
MIT. 42 claim-support items. Five runs on September 22. Hard 18: Jev 73.3%, GPT-5.4 70.0%, p = 1.000. A different project from fstandhartinger/jevbench.
mizchi/jev-playground
Fit table in docs/fit.md. 489 MOBA decisions, 0 illegal. One-line rule 16/16 against Jev at 6 to 27 wrong pages. No LICENSE file at the repo root.
2nd1st/Jevsus
MIT. True/false 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.