The full directory
All Jev AI model resources
Entries for TypeSafe AI's flagship System One model — official docs and SDKs alongside community clients, tools, and write-ups.
- Entries
- 3,919
- Categories
- 11
- Official
- 102
- Community
- 3,817
- Source sites
- 293
- Deep dives
- 3,919
By format
3919- Docs169
- Repos & sites3102
- Videos74
- Posts104
- Articles392
- Playgrounds78
Updated
Browse the Jev AI model by category
11 ways into the ecosystem. Each category page leads with its own shortlist of starting points.
Browse the Jev AI model by use case or industry
The same directory, sliced by what you are building and who you are building it for.
All 3919 Jev AI model resources
Search the full directory, or filter by category and tag.
Showing 1585–1632 of 3919 resources
CalOn is a unified 0.6B-parameter encoder that jointly encodes text and type definitions to extract entities, relations, and event triggers and arguments across datasets and schemas. Across 40 benchmarks it outperforms the evaluated baselines on most and leads all reported relation extraction benchmarks.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
An audit of request independence in seven open Jev-style checkpoints, using 276 three-passage states built from 92 TREC retrieval queries. Paired interventions show request expansion flips 9.1%/8.0% of Nimble's judgments and changes up to 21.4% of selection-or-abstention actions.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A study of whether Jev can serve inside reinforcement learning training rather than as a single-domain black box. Frozen Jev fills reference-policy, exploration-judge, and replay-rater roles, and training with it beats a standard RL learner across nine MiniGrid tasks and three Atari games.
Jev Practices & PatternsArticle#community#research#arxiv
Community
Reference-bound Visual Jev rewards verify whether requested attributes, actions, and relations hold for specified reference subjects in multi-subject image generation. Training a Qwen3.5-4B verifier as a GRPO reward raises a GPT-5.4 composite score from 41.78 to 52.50 on a 897-task MICo-Bench subset.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A study of using Jev inside Stockfish search and multi-teacher distillation. A bot with Jev's judgment distilled into a compact evaluator climbs above a 2200 Lichess bullet rating, and averaging Jev with Qwen3-32B labels beats spending the whole labeling budget on Qwen alone by 9.6 Elo.
Jev Practices & PatternsArticle#community#research#arxiv
Community
SanSi turns a pre-trained looped language model into a typed decision model that revises its hidden state across loops without generating tokens. On 10,027 test decisions from 59 sources it reaches 72.0% accuracy, 13.5 points above a same-shaped non-looped model trained with the same recipe.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
JevCascade allocates inference-time compute for prefill-only decision models using readout stability: when only the candidate menu changes, post-intervention accuracy is already determined by the cached first-pass distribution. Menu-only interventions are predicted to within 4.2 points across six model families.
Jev Practices & PatternsArticle#community#research#arxiv
Community
An analysis of open-weight typed decision models that finds their probabilities mostly follow option labels rather than written definitions. Deleting every definition leaves laya-td accuracy unchanged (0.8559 versus 0.8487), while renaming options to A and B raises accuracy by 0.1511.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A student-guided distillation pipeline for routing across 60 LLM task categories: a ModernBERT student retrieves top-k candidates and a DeBERTa-v3 zero-shot NLI teacher reranks only those. The best checkpoint reaches 77.5% teacher agreement, with Coverage@16 of 91-100%.
Jev Practices & PatternsArticle#community#research#arxiv
Community
LLM-as-Jev asks how far off-the-shelf LLMs already go as Jev-style decision models and when fine-tuning is needed. Frozen Qwen3.5-4B and Qwen3-0.6B match or exceed existing training-free readouts on the same 4B backbone, and Thinking Jev adds budgeted test-time reasoning.
Jev Practices & PatternsArticle#community#research#arxiv
Community
SBERT2S1 converts biomedical Sentence-Transformers encoders into typed decision models with bi-encoder cross-head and prior-fused residual variants, alongside BIODECIDE and MEDLINE-S1. Retrieval training helps the prior-fused head in 10 of 15 comparisons after fine-tuning.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
4GitHub stars
An analysis of ordinal-scale bias in JEV 1.13 and three open KEV models: across 36 ordinal datasets, final decisions use only 67-76% of the effective gold support versus 87-102% on four nominal tasks. BA-LoRA post-training raises utilization from about 47% to 86%.
Jev Practices & PatternsArticle#community#research#arxiv
Community
Glax147/jev_ordinal_scale_biaPython
Glax147/jev_ordinal_scale_biaPythonDyad augments a pretrained LLM with an environment-conditioned action encoder that scores candidate action descriptions to yield a distribution over typed actions. Training only the encoder helps across four unseen environments; joint training adds a 3.80% absolute gain on ALFWorld with a 9B model.
Jev Practices & PatternsArticle#community#research#arxiv
Community
Koa-decision: Fast and Consistent Structured Decision Making with Generative LLMs (arXiv 2609.36115)
Koa-decision reduces single-step decisions such as classification, endpointing, Boolean checks, and scoring to constrained generation with single-token outputs. On a production intent-routing benchmark it reaches 85.5% accuracy in about half a second, up to roughly 7.5x faster than frontier models.
Jev Practices & PatternsArticle#community#research#arxiv
Community
3GitHub stars
PACT turns contrastive training pairs into four loss terms for single-token typed-decision models. On a frozen 324-item holdout it matches the published recipe's accuracy (84.6% versus 85.2%) while cutting answer flips under relabelling from 13.8% to 9.8%.
Jev Practices & PatternsArticle#community#research#arxiv
JevSoup is a training-free framework that separates System One expert routing from System Two execution: Jev picks two LoRA experts through structured probabilities and their updates are merged with orthogonal projection. Across 14 PorTAL tasks it gains up to 1.19% task-macro accuracy over external baselines.
Jev Tools & CLIsArticle#community#research#arxiv
Community
Leowang980/JevSoupPython
Leowang980/JevSoupPythonAn evaluation of Jev and two open-weight typed decision models when option names and definitions conflict. With yes/no names, decision-flip rates are up to 70.4 percentage points higher than the 0/1 control, and mean AUC drops from 93.8% to a below-chance 23.2%.
Jev Practices & PatternsArticle#community#research#arxiv
Community
StartLux-Decision is a family of decision models from 0.8B to 27B parameters pairing single-pass Instant variants with AutoThink versions that reason only when uncertain. On the public subset of Decision Index 0.3 the 27B AutoThink model reaches 73.4.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
JevOut shows short background context additions can flip correct decisions in four systems including Jev across seven datasets. Within 64 accepted evaluations per decision, additions redirect 61.4%-73.2% of initially correct decisions toward a wrong option fixed in advance.
Jev Practices & PatternsArticle#community#research#arxiv
Community
xzx34/JevOutPython
xzx34/JevOutPythonA comparative study of system-one decision models across 15 medical datasets and 8,742 instances, measuring accuracy, decision stability, and abstention. Jev rivals leading models on text-driven tasks and is weaker on numerically driven ones, while most models still fail to recognize missing answers.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
An exploratory study of JevForest, a budgeted feature-acquisition policy that aggregates path-dependent proposals from bootstrapped trees and queries Jev for semantic answers. On small held-out samples it beats static and random orderings on AG News but not on TREC.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A system that joins coded and narrative views of 5,601,890 Texas crashes from 2017 to 2025 into population estimates: Kumo Tabular reads coded records, Jev reads narrative samples, and human judgments recalibrate the probabilities. Phone use appears in 15,074 injury crashes in narratives versus 7,340 in coded fields.
Jev Tools & CLIsArticle#community#research#arxiv
Community
pozapas/kumo-jev-crash-recordsPython
pozapas/kumo-jev-crash-recordsPythonA training-free framework with two implementations that turn emotion classification into Jev judgments: one decomposes tasks into atomic judgments, the other aggregates complementary judgment paths. Across eight datasets standard Jev averages 62.93% macro-F1 versus 67.28% for the strongest LLM baseline.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A controlled evaluation of Jev as a source-support verifier for GPT-4.1-mini calculation traces on financial reports. Holding operands and arithmetic fixed while swapping citations between same-number cells reveals wrong-role citations that pass and valid alternative citations that are withheld.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A semantic database system that answers filters, joins, classification, and ranking with fast typed decision models and escalates uncertain cases to LLMs. JEVDB-Flash posts the lowest latency on all 21 SemBench queries, and Semantic Bloom Filters remove 87.4% of join candidates on Shelob.
Jev Tools & CLIsArticle#community#research#arxiv
Community
A decision-native interface that represents each medical decision as a Choice, Noul, or Score over a runtime candidate set and returns a calibrated distribution over it. Against a matched generative baseline it cuts calibration error by an order of magnitude, with counting the one task where the baseline leads.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
lytang63/OmniMed-JevPython
lytang63/OmniMed-JevPythonA first empirical study of Jev on application classification for network flows, using ten packet-header features from CESNET-QUICEXT-25. Forty labeled examples raise accuracy from 9.80% to 28.42%, while tree ensembles trained on 8,000 records reach about 67-70%.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A population-scale surveillance study in which Jev reads every Texas police crash narrative from 2017 through 2025, 5,018,079 in total, under an eight-question schema. It estimates 5,467 crashes with a documented pregnancy, including 58 documenting fetal harm.
Jev Tools & CLIsArticle#community#research#arxiv
A report identifying an arithmetic-dependent rejection bottleneck in Jev: on paired arithmetic problems answer-present accuracy reaches 99% while correct rejection falls to 7%. A development-selected decision threshold lifts rejection from 7% to 79% while retaining 97% answer-present accuracy.
Jev Practices & PatternsArticle#community#research#arxiv
Community
7GitHub stars
A system that decomposes all three tasks of SemEval-2026 Task III Track A into typed Jev decisions aligned with 488 CPU-fitted coefficients. It reaches 1.0645 RMSE on valence-arousal regression and 52.09/44.06 continuous F1 on extraction, above fine-tuned Llama-3.3-70B and GPT-OSS-120B baselines.
Jev Practices & PatternsArticle#community#research#arxiv
Community
ZhangYiqun018/jev-dimabsaPython
ZhangYiqun018/jev-dimabsaPythonA vision-language navigation framework that replaces multimodal generation with visual compression and typed action selection: Action-Centric Visual Compression builds action representations, Discriminative Action-Semantic Memory keeps them distinct. On R2R-CE it reaches 27.0% SR and 22.4% SPL at 0.65 s per step.
Jev Models & ReimplementationsArticle#community#research#arxiv
Community
A diagnosis-guided repair agent whose CWE classification comes from Jev. Jev beats six open-weight models on every classification and ranking metric and runs at 6.27x lower median latency than GPT-5.6-Sol; with Jev diagnosing, repair lifts the detector-measured security pass rate from 63.5% to 70.7%.
Jev Tools & CLIsArticle#community#research#arxiv
Community
An exploratory offline evaluation of decision readouts for text-mediated video anomaly detection on 40 videos and 400 anchors from UCF-Crime and XD-Violence. Jev Noul reaches 75.99% average precision on XD captions but shows no corresponding UCF advantage.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
An evaluation of Jev 1.13 on four medical benchmarks: MetaMedQA, PubMedQA, DiagnosisArena-MCQ, and the NEJM Case Challenges, with GPT-6 Sol as reference. All 8,469 requests returned a valid answer; Jev matched GPT-6 Sol on PubMedQA but trailed on diagnostic cases.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A study posing speech-neuroprosthesis rescoring as a single typed Jev decision. On 978 held-out sentences from a participant with ALS, Jev reaches 7.5% word error versus 7.8% for OPT-6.7b and Qwen2.5-7B, at USD 0.07 per thousand sentences with no GPU needed.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
An evaluation of Jev and nine language models on ContractNLI with repeated requests that vary visible hypotheses, requested answers, and output order. Corrections and regressions can cancel so accuracy stays unchanged while decisions change; Jev is cheapest and fastest for the joint baseline.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
An evaluation of Jev as a low-cost judge of agreement with physician-written radiology reports. A single-question configuration reaches Kendall correlations of 0.573 on RadEvalX and 0.398 on RadEvalExpert, and one support question per statement keeps expert agreement while using 43-45% fewer input tokens.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
51GitHub stars
A StarCraft II controller combining fast JEV action selection with persistent GPT-6 planning that defeats the strongest non-cheating built-in AI, Lv7. The combined system wins four full games at Lv5-Lv7 at a mean estimated cost of USD 3.71 per game, with a median JEV response time of 0.422 s.
Jev Tools & CLIsArticle#community#research#arxiv
Community
sc2musa/Jev_StarPython
sc2musa/Jev_StarPythonAn evaluation of Jev as a semantic decision component in scientific workflows using a harness that assigns arithmetic to code. Across twelve model configurations, twenty source-grounded Choices, and ten scientific cases, Jev matched five configurations at complete semantic correctness.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A zero-shot video quality model that feeds encoding metadata, bitstream statistics, or pixel features to Jev. On 1,936 AV1, H.264, HEVC, and VP9 encodes, metadata-only reaches a Pearson correlation of 0.737 against VMAF, on par with ITU-T P.1204.1, and added features lift it to 0.824.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A system that codes Texas crash narratives as gated, typed Jev decisions returning probabilities and no text: a screen covered 499,500 narratives and 195,857 were coded with a 27-question schema. The typed model attains F1 of 0.908 against human labels, and recalibration cuts calibration error 3.3-fold.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A comparison of Jev-1.13.0 and two other decision models with LLMs for intent interpretation in 6G Open RAN. Jev meets the 1 s near-real-time budget on 99.8% of calls while two hosted LLMs meet it on 17.9% and 0%, with median interpretation taking 0.286 to 2.35 s.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
An integration of Jev's decision API into edge service orchestration that extracts four to eight bounded intent fields per request. Across 33 test conditions Jev cuts median decision latency by 22.7-64.5% versus the fastest LLM, and on a live admission path it keeps 0.91-0.95 of requests exact and on time.
Jev Tools & CLIsArticle#community#research#arxiv
Community
An evaluation of Jev on training-free human activity recognition using 1,800 class-balanced accelerometer windows from WISDM, UCI341, and PAMAP2. Its strongest representation reaches macro-F1 of 0.038, 0.118, and 0.089, versus 0.686 to 0.907 for the supervised models.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
A behavioral diagnosis of why System One decision models choose wrong actions in agent loops. On a frozen benchmark of agent decision states, most wrong actions follow a correct reading of the state, and the more accurate the system, the more of its errors take this form.
Jev Practices & PatternsArticle#community#research#arxiv
Community
A pre-registered evaluation of Jev as a memory-retirement gate deciding when a new utterance retires an old stored claim. On 500 labeled test pairs, the binary gate retires 39.2% of claims whose change is only suggested; a three-way contract cuts that to 15.5% at a 15.3-point detection cost.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
An instrumented spatial predator-prey framework for studying communication and collective dynamics in AI agents, separating the evolving world, request scheduling, controller adapters, and experiment records. A case study reports a signals-on minus signals-off rabbit-abundance difference of -0.0095.
Jev Tools & CLIsArticle#community#research#arxiv
Community
A paired, self-audited evaluation of an open-weight (Laya) and a hosted (Jev) System-1 decision model on 11 agent decision points: 7,283 base cases plus 6,640 robustness variants. Jev is significantly more accurate on 9 of 11 decision points, by 10.8 to 46.0 points.
Jev Benchmarks & EvaluationsArticle#community#research#arxiv
Community
David-DL-Space/sys1-evalJavaScript
David-DL-Space/sys1-evalJavaScriptStart here with the Jev AI model
Hand-picked entries the maintainers recommend first, from official primitives to community tools.
The Jev Day release post: Jev 1.14.0 under jev-preview plus three danger-zone features — composable images on both models, the box question type on 1.14 only, and opt-in 262k context.
Jev Guides & ArticlesArticle#official#release#vision
Official
The box question type: one bounding box per question, returned as four edges that are fractions of the image, plus confidence, with no criteria field and no not-found value.
Jev Practices & PatternsDocs#official#docs#vision
Official
LiteLLM's Day 0 post for TypeSafe Jev 1.14: the proxy adds box questions, image input and long-context mode on both POST /v1/systemone and POST /v1/decisions, forwarding TypeSafe-Danger-Zone headers only to TypeSafe deployments.
Jev Integrations & AdaptersArticle#community#litellm#gateway
Community
Newest Jev AI model resources
The latest additions to the directory, newest first.
Do Multimodal Models Decide Consistently on Equivalent Optical Spectra?Jev Benchmarks & Evaluations
@harlan-zw/jevJev Repos & SDKs
Jev-AD: Safe Language-Guided Autonomous Driving via a Language World ModelJev Demos & Experiments
Jev GUI Training PreviewJev Models & Reimplementations
- jev-playgroundJev Demos & Experiments
- JeviewJev Tools & CLIs
From the Jev community
Watch walkthroughs, read the threads, and see what builders are posting.
YouTube: KodeKloud's three-minute primer: what a System One model decides, and why it answers in milliseconds.
YouTube: Caleb Writes Code explains what Jev is and how to call it, in seven minutes.
YouTube: Greg Isenberg and Ryan Vogel on the businesses Jev unlocks, and a path in through the Vercel AI Gateway.
YouTube: CJ from Syntax demos browser use, classification, code review, a model router, and a chat bot with no LLM.
YouTube: Ryan Vogel runs Jev on 100 and then 1,000 emails: category, priority, spam, and reply predictions.
YouTube: AICodeKing tests support routing, refund detection, prompt-injection resistance, and browser automation.
Hacker News: Show HN: Ad Spend Guard gates budget before the auction, not the creative after.
Show HN: Gating ad budget with a decision modelRead the thread on Hacker NewsReddit: Reranking rebooking options with a decision model, ranked with reasons.
Hacker News: Pick Verify: a single decision at the warehouse pick station catches confusable picks.
Show HN: Verifying robot picks with a decision modelRead the thread on Hacker NewsHacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Software Bombe and live Enigma demo where Jev ranks cribs and judges plaintext.
ag_212 pointsShow HN: Jev as the Bletchley Park Analyst: A Software Bombe on EnigmaRead the thread on Hacker NewsHacker News: Show HN for the decide CLI: brew install, file context, and multi-question shell scripts.
vsekhar2 pointsShow HN: Decide – Jev decisions in the shell, scripts, and agent skillsRead the thread on Hacker NewsHow entries are chosen
Every entry is reviewed against the same bar, whether it comes from TypeSafe or the community.
- Public and working — no paywalls, sign-up walls, or dead links.
- Concrete — it has to be about Jev and the System One API, not generic LLM content.
- Useful today — something a developer can learn from or use right now.
- Distinct — if a similar entry exists, the description has to say what makes this one different: a different runtime, a cheaper path, a measured result, or a narrower focus.
- Documented — repos state their license, and docs are in English or carry an English summary.
Spam, affiliate links, and thin AI-generated content are rejected. New entries are expected to ship with a deep dive: two to five highlights, caveats, a quickstart snippet where one helps, and a verification date. Star counts are refreshed when an entry is touched.
Common questions
Short answers about Jev and how this directory works.
What is the Jev AI model?
Jev is TypeSafe AI's first System One model — an AI model that takes a state plus typed questions and returns one structured answer per question: a choice, an ordered score, or the probability of yes, with calibrated confidence on Choice and Score answers. Code can branch on the result without parsing generated text.
What is the difference between Choice, Score, and Noul?
They are the three question primitives. Choice picks one of up to 255 options, Score rates on 2 to 10 ordered levels, and Noul answers yes or no. Choice and Score answers carry a confidence between 0 and 1; a Noul answer is the probability of yes.
Which Jev SDKs and client libraries are available?
TypeSafe maintains official JavaScript/TypeScript and Python SDKs plus a Python adapter. The community has built clients for Go, .NET, PHP, Laravel, Rust, Elixir, Ruby, and Scala, along with framework integrations such as DSPy, LlamaIndex, and Home Assistant. All of them are listed in the Repos & SDKs category.
Is the directory free, and is it affiliated with TypeSafe AI?
The directory is free to browse: every entry is reviewed, and spam, affiliate links, and sign-up walls are rejected. JevDirectory.org is community-run and not affiliated with TypeSafe AI — Jev and System One are TypeSafe AI trademarks.
How often is the directory updated?
Entries are added as new resources appear and re-checked when they are touched. The directory currently holds 3919 entries from 293 source sites, last updated Nov 6, 2026.
What is a verified deep dive?
Most entries include a deep dive: two to five highlight bullets someone would only know from using the resource, a caveats note about limits and requirements, an optional quickstart snippet, and the date the link was last checked. 3919 of the 3919 entries carry one today.
For the canonical details, read the official TypeSafe docs and the API reference.