# Decision models in an independent AI registry > DecisionModels.ai is an independent registry of 60 decision model entries, with source links and labelled claims alongside published prices. Updated 2026-11-04 (Australia/Adelaide) ## AI for decisions. **Decision models** return structured answers, such as categories, scores or yes/no probabilities, for games, websites and AI agents. [Jev](https://decisionmodels.ai/models/jev) returns these without generating text; some alternatives generate an option letter. [Compare decision models]() [LATEST: hs-decide-1-flash: open, 3 Nov]() ## Releases since Jev launched on 15 September TypeSafe AI released [Jev](https://decisionmodels.ai/models/jev) and named the category. Every model below links to its model page. 11,206 public GitHub repos matching "jev", created 15 to 30 Sep. Checked 03 Nov 2026 · 27 created 11 to 14 Sep. [Source]() 44 registry entries follow Jev on this timeline. 252 Jev-class systems measured on JevBench, an independent leaderboard. v1.6.37, 21 Oct 2026. [Source]() - [Jev]() · TypeSafe AI · Released: 2026-09-15 - [Laya]() · Convai Innovations · Released: 2026-09-18 - [Nimble]() · Bespoke Labs · Released: 2026-09-18 - [Decision 1.0]() · vLLM Semantic Router · Released: 2026-09-22 - [Tev1-4B]() · Together AI · Released: 2026-09-23 - [Span-01]() · Respan · Released: 2026-09-24 - [GLiNER2.5-Decide]() · Fastino · Released: 2026-09-24 - [Intern-Decision]() · InternLM · First Hugging Face upload: 2026-09-26 - [Solar Decide]() · Upstage · OpenRouter announcement: 2026-09-28 - [Ollama v0.35]() · Ollama · Released: 2026-09-28 - [d1]() · Liquid AI · Released: 2026-09-29 - [Mercury Decide]() · Inception · Released: 2026-09-30 - [Kev]() · Jared Palmer · Released: 2026-10-01 - [DecisionTune 1.0]() · DecisionTune · Hugging Face repository upload: 2026-10-04 - [Valen]() · Valen Team · First Hugging Face weights upload: 2026-10-05 - [Decisions API]() · OpenAI · Public beta launch: 2026-10-06 - [Decider]() · Perplexity · Released: 2026-10-06 - [H2O-Lightning-4B]() · H2O.ai · Released: 2026-10-07 - [KnowLine]() · PelaAI · First Hugging Face family release: 2026-10-07 - [Snowflake Decision]() · Snowflake · Released: 2026-10-08 - [Microsoft-Decision-1]() · Microsoft · Released: 2026-10-09 - [Clef-omni]() · Cloudflare · Released: 2026-10-09 - [Arc-decide]() · Arc53 (DocsGPT) · Hugging Face v0.2 upload: 2026-10-11 - [Granite Switch 4.1]() · IBM · Released: 2026-10-15 - [TYGER]() · NineNineSix · Released: 2026-10-17 - [Clairic]() · Cloudflare · Released: 2026-10-19 - [GLiDE]() · Fastino · Weights repository upload: 2026-10-19 - [Zoom Decisions API]() · Zoom · Released: 2026-10-19 - [Qev 0.8B]() · labqoat · Hugging Face weights upload: 2026-10-20 - [Mammoth 1]() · Baseten (Base Labs) · Released: 2026-10-21 - [Talos]() · Tractian · Repository and weights upload: 2026-10-22 - [Decido]() · decido-ai · Weights upload: 2026-10-22 - [LingDT-Decision-1.0]() · Ant Digital Technologies (蚂蚁数科) · Released: 2026-10-22 - [MiniCPM-OCR Jev]() · OpenBMB · First Hugging Face upload: 2026-10-23 - [Fovea-4B]() · Sivasuthan · Weights upload: 2026-10-24 - [Parallel Decisions API]() · Parallel · Released: 2026-10-25 - [homi]() · egma-ai · Hugging Face weights upload: 2026-10-27 - [UltraBERT Curation]() · Daniel van Strien · Released: 2026-10-28 - [Nimbus]() · Crele · Hugging Face weights upload: 2026-10-30 - [Ru-Jev]() · besktech (beskvladimir-create) · First Hugging Face repository creation: 2026-11-01 - [Hume Audio Decisions API]() · Hume AI · Released: 2026-11-02 - [Smaug-Decisions]() · Abacus.AI · Maker blog announcement: 2026-11-02 - [hs-decide-1-flash]() · Vectorize (Hindsight) · Released: 2026-11-03 - [REAX 1.3]() · REAX · Released: 2026-11-03 - [K-Decision-9B]() · ONYULTEA · First Hugging Face upload: 2026-11-03 ## What happens in one call A message or an agent's history can be input. You can also send a game state or a web page. State: "My card was charged twice." For a Choice question, you supply the allowed options and receive probabilities for them. which team? billing · technical · sales Use the probabilities to decide which answers to act on and which to send to a person or a bigger model. Illustrative odds: billing .94 · technical .04 · sales .02. route("billing") The odds above illustrate a Choice question. [Jev](https://decisionmodels.ai/models/jev), and the models built to its API, also take **Noul** questions, which return a yes/no probability; the name comes from the Bernoulli distribution. A **Score** question returns a value along your defined scale, which can fall between the numeric levels you supply. [source](https://simonwillison.net/2026/Sep/21/jev/). ### How generative LLMs compare A generative LLM can return structured output that your code reads. It generates its answer one token at a time. [Jev](https://decisionmodels.ai/models/jev) returns structured answers without generating text. Together reports that [Tev1](https://decisionmodels.ai/models/tev1) keeps a text-generation head and is trained to return one option letter. [Simon Willison](https://simonwillison.net/2026/Sep/21/jev/) [Together](https://huggingface.co/togethercomputer/Tev1-4B-experimental) ### Where the name comes from TypeSafe calls [Jev](https://decisionmodels.ai/models/jev) a *System One model*, after Daniel Kahneman's split between fast, intuitive thinking and slow, deliberate reasoning. Liquid AI uses "decision models" for [d1](https://decisionmodels.ai/models/d1). [TypeSafe](https://typesafe.ai/blog/introducing-system-one-models-and-jev) [Liquid](https://docs.liquid.ai/lfm/models/decision-models) ### Is it new? The interface isn't. Scoring declared labels in one pass goes back years, from zero-shot classifiers to GLiNER. What is new is a hosted model sold for it, with calibrated probabilities as its stated training goal, at this price and speed. [The lineage](https://decisionmodels.ai/#lineage) ## What people are using decision models for The newest builds we've found. Each card links to its use case, with a source for the example. ### Comment analysis [Comment analysis]() What topic does this comment discuss? Options: combat · technical performance · other topics Example: [pbz-comment-insight]() · [Source]() ### Agent memory [Agent memory]() Is this fact new, or already known? Options: new · already known Example: [Hindsight agent memory]() · [Source]() ### Alert paging [Alert paging]() Should the on-call engineer be paged now? Options: page · do not page Example: Coroot paging benchmark · [Source]() ### Aerial imagery [Aerial imagery]() How many RVs are in this tile? Options: 0 · 1 · 2 · 3 · 4 to 5 · 6 to 8 · 9 to 12 · 13 to 20 · 21 to 35 · 36 or more Example: [clef-playa counts aerial tiles with Clef]() · [Source]() ### OCR checks [OCR checks]() Score whether a candidate OCR reading matches its document region. Options: 0 to 1 correctness score Example: [MiniCPM-OCR Jev checks text, formulas and tables]() · [Source]() ### Voice quality checks [Voice quality checks]() Does the agent talk over the caller? Options: yes · no Example: [Hume Audio Decisions API docs]() · [Source]() ### Live web facts [Live web facts]() Does this vendor support SAML SSO and EU data residency? Options: yes · no Example: [Parallel Decisions API benchmark post]() · [Source]() ### Document retrieval [Document retrieval]() Does the answer to this question lie under this section? Options: yes · no Example: [Agastya]() · [Source]() ### Context hygiene [Context hygiene]() Has the conversation moved to a new topic? Options: same · subtopic · new\_topic Example: [topic-gate]() · [Source]() ### Note tagging [Note tagging]() Is this note about a trip? Options: yes · no Example: [Laya Notes]() · [Source]() ### Live audio [Live audio]() Is music playing? Options: yes · no · cant\_tell Example: [Nimbus]() · [Source]() ### Mobile apps [Mobile apps]() Could executing this action cause data loss, a charge, a logout, an account or permission change, a message sent to another person, or another irreversible side effect? Options: yes · no Example: [jev-droid]() · [Source]() ### Agent skills [Agent skills]() Does this output follow this rule from the skill? Options: yes · no Example: [skill-tester]() · [Source]() ### Reasoning effort [Reasoning effort]() How much reasoning effort does this task need? Options: low · medium · high Example: [jev-router for Claude Code]() · [Source]() ### Group chats [Group chats]() Should the bot reply now? Options: true · false Example: [JEV Damocles for AstrBot]() · [Source]() ### Code search [Code search]() Does this declaration implement the behaviour in the query? Options: yes · no Example: [layagrep]() · [Source]() ### Earnings calls [Earnings calls]() Does this claim from the call contradict the reported figures? Options: yes · no Example: [narrative-audit-jev]() · [Source]() ### Video editing [Video editing]() Is there room in this shot for one more person? Options: yes · no Example: [Cameo for ComfyUI]() · [Source]() ### Filing loose files [Filing loose files]() Which project does this file belong to? Options: Podcast · Client site · Home admin · None of these Example: [Unorphan]() · [Source]() ### Content moderation [Content moderation]() Is the user trying to find, track or obtain private information about a specific person? Options: yes · no Example: [Exposure-aware routing]() · [Source]() ### Disk cleanup [Disk cleanup]() Does this folder contain only caches, temporary files, crash dumps or leftover downloads that apps recreate? Options: yes · no Example: [SpaceSniffer, Jev edition]() · [Source]() ### Video footage [Video footage]() Is the mixing step happening at this time? Options: Present · Absent · Unclear Example: [Jev]() Video Lab · [Source]() ### Office documents [Office documents]() Do the supplied slides include a [pricing]() table? Options: true · false Example: [idec for local Ollama]() · [Source]() ### Support tickets [Support tickets]() Which team should handle this support ticket? Options: Billing · Technical · Account · Sales Example: [laya-pet ticket router]() · [Source]() ### Marketing ops [Marketing ops]() Is this reply interested? Options: yes · no Example: [Jev for Marketers guide]() · [Source]() ### Issue backlogs [Issue backlogs]() Is this issue from someone blocked from upgrading? Options: yes · no · unknown Example: [Better Backlog for GitHub issues]() · [Source]() ### Answering the agent [Answering the agent]() Which of the options the agent offered should it go with? Options: Option 1 · Option 2 · Option 3 Example: [opencode-auto-question plugin]() · [Source]() ### Tool loading [Tool loading]() Does this request need the memory tools? Options: yes · no Example: [pi-jev-lazy-load for Pi]() · [Source]() ### IT compliance [IT compliance]() Which remediation path should resolve this failing custom setting? Options: Configure devices · Ask users · Fix policy · Review Example: [intune-why for Microsoft Intune]() · [Source]() ### Marketplace listings [Marketplace listings]() Does this listing match what I'm looking for? Options: yes · no Example: [Jev Content Filter for classified ads]() · [Source]() ### UI testing [UI testing]() Which element on this page does the step mean? Options: Checkout button · Cart icon · First name field · Continue button Example: [Decisium, plain-English UI tests]() · [Source]() ### Order matching [Order matching]() Which catalogue item is "5# limes"? Options: Limes, 5 lb · Limes, case · Lemons, 5 lb · none of these Example: [order-matcher for restaurant orders]() · [Source]() ### Terms of service [Terms of service]() How does this clause treat users in disputes? Options: not addressed · standard or protective · somewhat stricter · abusive Example: [Jev ToS Checker for Chrome]() · [Source]() ### CI failures [CI failures]() What caused this failed test? Options: environment or timeout · flaky assertion · genuine regression · API contract mismatch Example: [Jev CI triage]() · [Source]() ### Source screening [Source screening]() Does this news item carry PRC state framing? Options: yes · no Example: [West Philippine Sea Watch]() · [Source]() ### Data compliance [Data compliance]() Does this column hold personal data? Options: yes · no Example: [jev-pii, PII discovery in DuckDB]() · [Source]() ### Notifications [Notifications]() What kind of attention does this notification need? Options: action required · ephemeral interrupt · persistent status · review later · hide Example: [QuietFlow for Android]() · [Source]() ### Idea vetting [Idea vetting]() Does this existing repo already do the same job as the idea? Options: yes · no Example: [buildornot, a verdict on your side project]() · [Source]() ### Simulations [Simulations]() What does this villager do for the next two hours? Options: eat · rest · work · explore Example: [The Village, run on Jev]() · [Source]() ### Tutoring [Tutoring]() Can this question be answered from the textbook chapter? Options: yes · no Example: [Textbook Voice Tutor, Grade 2]() · [Source]() ### Occupation coding [Occupation coding]() Which major group does this job ad belong to? Options: Professionals · Associate professionals and technicians · Clerical support workers · Service and sales workers Example: [SSOC 2024 Decision Trail]() · [Source]() ### Live calls [Live calls]() Has the other person answered this question yet? Options: yes · no Example: [Call Coach, a live interview checklist]() · [Source]() ### Product catalogues [Product catalogues]() Does this product's text say it contains chicken? Options: supported · contradicted · unknown · conflicting Example: [Petsona, evidence-backed pet product tags]() · [Source]() ### Paper screening [Paper screening]() Is this new paper about humanoid whole-body control? Options: yes · no Example: [arXiv Jev Digest, a daily screened reading list]() · [Source]() ### Medicine [Medicine]() Which tissue type is in this slide tile? Options: tumour · normal · inflamed · none of these Example: [PathJev, decisions on pathology slides]() · [Source]() ### Industrial control [Industrial control]() Should this agent's request reach the controller? Options: allow · cache · defer · deny Example: [Talos gate for AI agents on factory networks]() · [Source]() ### Email triage [Email triage]() Is this email for you? Options: for you · to review · junk Example: [Mail Triage, on-device with Laya]() · [Source]() ### Feed filtering [Feed filtering]() Does this post match my interests? Options: show · hide Example: [LinkedIn Focus Filter, in-browser with Kev]() · [Source]() ### SQL queries [SQL queries]() Does this record's text match the question? Options: yes · no Example: [Laya inside DuckDB]() · [Source]() ### Coding assistants [Coding assistants]() How big is this request? Options: question · small change · big change · too vague Example: [ClauJev, Jev checking Claude Code]() · [Source]() ### Voice agents [Voice agents]() Should the agent talk now? Options: speak · keep listening · yield · mm-hm Example: [TYGER full-duplex voice agent]() · [Source]() ### Visual inspection [Visual inspection]() Does this pipe pass inspection? Options: pass · defect · recheck Example: [OpenCV + Laya inspection workbench]() · [Source]() ### Agent guardrails [Agent guardrails]() Should this bash command run? Options: allow · block Example: [guardrails-md for coding agents]() · [Source]() ### Forecasting [Forecasting]() Will this happen by the deadline? Options: yes · no Example: [Three decision models vs Polymarket and Kalshi]() · [Source]() ### Games [Games]() What should the player do next? Options: left · right · jump · wait Example: [Mario, played from emulator state]() · [Source]() ### Browser agents [Browser agents]() Which element on this page gets clicked? Options: search box · results · next page Example: [browser-use: flights in 7.1 s]() · [Source]() ### Computer use [Computer use]() What's the next action on screen? Options: click · type · scroll · done Example: [typesafe-computer-use]() · [Source]() ### Search ranking [Search ranking]() How relevant is this result to the query? Options: a score on a scale Example: Simon Willison on reranking with [Jev]() · [Source]() ### Trading [Trading]() What do we do in this block? Options: buy · sell · hold Example: [jev-trader, one decision per block]() · [Source]() ### Robotics [Robotics]() Where does the drone go next? Options: forward · left · right · climb Example: [Camera-input drone in simulation]() · [Source]() ### Code and security [Code and security]() Is this commit a security fix? Options: yes · no Example: [commit-miner]() · [Source]() ### Cleaning up pages [Cleaning up pages]() Is this element clutter? Options: yes · no Example: [unclutter browser extension]() · [Source]() ## Every model we know of, with sources Hosted models, open weights you can run yourself, and the community projects around them. Tap any entry for what it is, what its maker claims, and where that claim comes from. Claims are marked **VENDOR-REPORTED** until someone independent reproduces them. Inclusion is not endorsement, and nobody pays to be listed. Something missing or wrong? [if@decisionmodels.ai](mailto:if@decisionmodels.ai) ### Jev [Jev]() By TypeSafe AI · Released: 2026-09-15 · HOSTED The model that named the category. A hosted API: send your app's state plus typed questions (Choice, Score or Noul) and get a probability back for each answer. It generates no text, and its weights are not public. #### Facts - Price: $0.042 per million input tokens; output free · [Compare prices]() - Latency: 70–500 ms end to end (TypeSafe-reported) - Version: jev-1.13 is the stable release; jev-1.14 has been the preview build since 14 Oct 2026 - Context: 32,000 tokens (per OpenRouter's guide) - Access: Launched in early access; also via OpenRouter #### Claims and sources - **VENDOR-REPORTED**: 40x–200x faster than frontier LLMs on comparable tasks [source]() - **VENDOR-REPORTED**: Trained with Reinforcement Learning for Calibrated Decisions (RLCD); no technical paper published as of 2 Oct 2026 [source]() - **INDEPENDENT**: Leads [JevBench]()'s capability ranking of Jev-class systems (80.0) on 1 Oct 2026 [source]() - **VENDOR-REPORTED**: No public weights: TypeSafe's own Hugging Face account listed zero public models when checked on 20 Oct 2026 [source]() - **VENDOR-REPORTED**: The 14 Oct 2026 update switched on experimental image input and opt-in long context for [Jev]() 1.13 and 1.14 (1.13 was "multimodal all along"); bounding-box answers need 1.14 [source]() - **INDEPENDENT**: Tops Artificial Analysis's Decision Index at launch with 58, ahead of [GLiDE]() and [Microsoft-Decision-1]() on 54 (13 Oct 2026) [source]() #### Links - [Source: Announcement]() - [Source: Jev 1.14]() - [Source: Docs]() - [Source: On OpenRouter]() ### Decisions API [Decisions API]() By OpenAI · Public beta launch: 2026-10-06 · HOSTED OpenAI's [Decisions API]() uses GPT-6 Luna to return typed answers for classification, routing and agent actions. It launched in public beta on 6 Oct 2026 and became generally available on 22 Oct. Developers supply the questions and allowed answers; the endpoint returns probabilities, choices or rubric scores. #### Facts - Model: gpt-6-luna only - Access: Generally available since 22 Oct 2026; public beta launched 6 Oct; native GPT-Live integration since 15 Oct - Price: $0.10 per million uncached input tokens; output free · [Compare prices]() - Inputs: Text, images, JSON and conversation history with tool calls and outputs - Questions: Up to 400 per request, increased from 200 in public beta - Endpoint: POST /v1/decisions #### Claims and sources #### Links - [Source: Decisions guide]() - [Source: API changelog]() ### d1 [d1]() By Liquid AI · Released: 2026-09-29 · HOSTED Liquid AI's decision model family. The hosted [d1]() takes the same three question types as [Jev]() (Noul, Choice and Score). On 7 Oct Liquid released two small open siblings you can run on a device, and on 15 Oct a bigger hosted one, d1-ultra, joined OpenRouter. On 27 Oct [d1]().1, the successor to [d1](), joined OpenRouter at the same price, with -latest aliases that follow each family's newest version. On 27 Oct Liquid announced two more open models, d1-2B (text, pictures and video) and d1-Omni-2B (adds audio), built on LFM3; their Hugging Face repositories went up on 24 Oct. #### Facts - Hosted: [d1]().1, [d1]() and d1-ultra, via Liquid's API and OpenRouter - Open weights: d1-3B and d1-omni-600M (7 Oct), then d1-2B and d1-Omni-2B (announced 27 Oct), under the LFM Open License v1.0 - Price: OpenRouter lists [d1]() and [d1]().1 at $0.02 and d1-ultra at $0.042 per million input tokens, output free (as of 28 Oct 2026) · [Compare prices]() #### Claims and sources - **VENDOR-REPORTED**: Liquid says [d1]() is the first model to beat [Jev]() on the Jev Decision Index, a community index hosted on Hugging Face. We could not find a [d1]() result in the index's public data (release 0.2.1, generated 28 Sep) [source]() - **VENDOR-REPORTED**: d1-3B answers a question in 16 ms on an NVIDIA Jetson AGX Thor [source]() #### Links - [Source: Docs]() - [Source: Announcement]() - [Source: Decision Index]() - [Source: Open d1 release]() - [Source: d1.1 on OpenRouter]() - [Source: d1-2B]() ### Mercury Decide [Mercury Decide]() By Inception · Released: 2026-09-30 · HOSTED Inception's decision model, served as a System One endpoint: send state and typed questions, get a choice, a score or a yes/no answer, each with a probability. #### Facts - Price: $0.02 per million input tokens at a 50% discount, output free; a free tier remains (OpenRouter, as of 22 Oct 2026) · [Compare prices]() - Context: 66K tokens (OpenRouter) - Access: OpenRouter #### Claims and sources - **VENDOR-REPORTED**: Up to 14 decisions per second [source]() - **VENDOR-REPORTED**: The most intelligent decision model on OpenRouter, on [JevBench]() v1.4 (the board was on v1.5.4 as of 2 Oct 2026) [source]() - **INDEPENDENT**: On Vals AI's LegalBench sample, 0.869 to [Jev]()'s 0.740 accuracy; on claim verification its router stayed inside a 1% error budget where [Jev]()'s went over [source]() #### Links - [Source: OpenRouter]() - [Source: Announcement]() ### Solar Decide [Solar Decide]() By Upstage · OpenRouter announcement: 2026-09-28 · HOSTED Upstage's decision endpoint, available through OpenRouter. The date is when OpenRouter announced it; we have not found an announcement from Upstage itself. A faster sibling, [Solar Decide]() Flash, joined on 8 Oct 2026 with a 524K context. #### Facts - Price: OpenRouter lists $0.05 per million input tokens, output $0 (as of 2 Oct 2026); the launch post described a limited-time 50% discount · [Compare prices]() - Access: OpenRouter #### Claims and sources #### Links - [Source: OpenRouter]() - [Source: OpenRouter announcement]() - [Source: Solar Decide Flash]() ### Span-01 [Span-01]() By Respan · Released: 2026-09-24 · HOSTED A one-pass classifier for agent traces. For each behaviour you define, it returns the probability that it is present, absent or not observable, without generating text. Its successor, Span-02 (15 Oct 2026), reads images too, can return a number directly, and ships as Apache-2.0 open weights. #### Facts - Price: $0.02 per million input tokens; output free · [Compare prices]() - Free tier: [Span-01]() Lite - Maker: Respan, a Keywords AI product #### Claims and sources - **VENDOR-REPORTED**: Beats GPT-6 Luna and [Jev]() 1.13 on Respan's own behaviour-detection benchmark [source]() #### Links - [Source: Announcement]() - [Source: Span-02]() ### Decider 1 [Decider 1]() By meraGPT · HOSTED A hosted decision model served only at /v1/systemone, the request format TypeSafe's SDK speaks. Its site doesn't say who runs it or when it launched. #### Facts - Price: $0.03 per million input tokens; output $0 · [Compare prices]() - Endpoint: /v1/systemone #### Claims and sources - **VENDOR-REPORTED**: 0.768 accuracy against [Jev]()'s 0.727 on typed-decision benchmarks [source]() #### Links - [Source: meragpt.com]() ### Microsoft-Decision-1 [Microsoft-Decision-1]() By Microsoft · Released: 2026-10-09 · HOSTED Microsoft's decision-scoring model for routing, classification, prioritisation, verification and workflow control. Post-trained from Qwen3.5-9B, available in Microsoft Foundry and on OpenRouter. #### Facts - Price: $0.042 per million input tokens; output free · [Compare prices]() - Base: Qwen3.5-9B - Access: Microsoft Foundry, OpenRouter #### Claims and sources - **INDEPENDENT**: Ties [GLiDE]() for second on Artificial Analysis's Decision Index at launch, on 54 (13 Oct 2026) [source]() #### Links - [Source: Announcement]() ### Clef-omni [Clef-omni]() By Cloudflare · Released: 2026-10-09 · HOSTED The [Clef]() family on Workers AI answers typed questions in one pass. [Clef-omni]() adds audio and video, so it can decide about what is seen and heard at the same time. The same update cut Clef-flash to $0.038 per million input tokens. #### Facts - Size: 30B mixture of experts, 3B active - Price: $0.150 per million input tokens · [Compare prices]() - Access: Cloudflare Workers AI; Clef-flash weights are on Hugging Face #### Claims and sources #### Links - [Source: Cloudflare changelog]() ### Snowflake Decision [Snowflake Decision]() By Snowflake · Released: 2026-10-08 · PREVIEW A decision model you call from SQL. It classifies, scores, filters and routes rows inside Snowflake through AI\_COMPLETE. A lower-latency Decision Fast, with REST access, followed on 19 Oct. #### Facts - Access: Private preview in Cortex AI Functions - Limit: Up to 32 options per question #### Claims and sources - **VENDOR-REPORTED**: Scored 57.63 on a subset of the Hugging Face decision benchmark (the 32-option limit rules out the full suite) [source]() #### Links - [Source: Announcement]() - [Source: Decision Fast]() ### Celeris-1-decision [Celeris-1-decision]() By Celeris · Released: 2026-10-08 · HOSTED A closed decision model for text, structured data and images that can give a one-sentence reason alongside each probability. #### Facts - Price: $0.04 per million input tokens; output free · [Compare prices]() #### Claims and sources - **VENDOR-REPORTED**: 80.0% on jev-bench at a 67 ms median response, ahead of [Jev]() 1.13 and [Mercury Decide]() on its own run [source]() #### Links - [Source: Announcement]() ### SubQ Snap [SubQ Snap]() By Subquadratic · Released: 2026-10-12 · PREVIEW A long-context decision model for text and images, with a 384K-token window. Version 1.1 (22 Oct 2026) cuts the published price to $0.04 per million tokens and is also available on Baseten. #### Facts - Price: $0.04 per million input tokens; output free · [Compare prices]() - Access: Public preview on Subquadratic; also served on Baseten - API model: subq-snap-1.1 #### Claims and sources - **VENDOR-REPORTED**: Subquadratic's 22 Oct 2026 release post quotes an Artificial Analysis Decision Index score of 65 for [SubQ Snap]() 1.1, tied for #1. This figure is quoted by the vendor. [source]() #### Links - [Source: Announcement]() - [Source: Version 1.1 release]() - [Source: Baseten]() - [Source: API models, status and pricing]() - [Source: Baseten launch, 02 Nov 2026]() ### Classify-1 [Classify-1]() By ScaleDown · OpenRouter listing: 2026-10-13 · HOSTED Scores text or a whole document against your own choice, yes/no and ordered-scale questions. Request-compatible with [Jev]()'s [Decisions API](). The date is OpenRouter's listing. #### Facts - Context: 1M tokens (OpenRouter) - Access: OpenRouter, ScaleDown API #### Claims and sources #### Links - [Source: Docs]() - [Source: OpenRouter]() ### Clairic [Clairic]() By Cloudflare · Released: 2026-10-19 · PREVIEW A new branch of the family: a System One extraction model. Where a decision model picks from answers you list, [Clairic]() extracts, computes, estimates or forecasts values from unstructured input and returns them as typed answers: an invoice total, a delivery date, a repair cost, a 24-hour traffic forecast. It does this quickly, without an LLM writing prose. #### Facts - Access: Workers AI as @cf/cloudflare/clairic; back on OpenRouter as cloudflare/clairic when checked 4 Nov 2026 - Weights: Cloudflare says open weights are coming soon (as of 19 Oct 2026) - Price: $0.10 per million input tokens on OpenRouter; output free · [Compare prices]() - Inputs: Text and images - Context: 65,536 tokens on OpenRouter #### Claims and sources - **VENDOR-REPORTED**: Higher accuracy, lower latency and lower cost than traditional LLMs at extraction [source]() #### Links - [Source: Announcement]() - [Source: OpenRouter listing]() - [Source: OpenRouter decisions inventory (34 entries, checked 4 Nov 2026)]() ### Parallel Decisions API [Parallel Decisions API]() By Parallel · Released: 2026-10-25 · HOSTED A decision engine that searches the live web before it answers. You send typed questions about a company, product or policy, and one call returns typed answers with probabilities (or a refusal) plus a shared list of the web pages it used, so it can answer things a model only knows from training data cannot, such as whether a company has an open job right now. #### Facts - Access: Hosted API, with a playground; 5,000 free requests to start - Price: $1 per 1,000 calls (as of 25 Oct 2026) · [Compare prices]() - Grounding: Searches the live web per request; citations cover the whole call, not each answer - Latency: VENDOR-REPORTED: 0.56 seconds median on PWS-Decisions, measured by Parallel on 29 Oct 2026 (Pacific). #### Claims and sources - **VENDOR-REPORTED**: Parallel says it costs 10x less than a standalone model [source]() - **VENDOR-REPORTED**: Parallel reports 76.22% on its own 300-question PWS-Decisions set with live web search, against Perplexity [Decider]() v1.2 at 50.89%, OpenAI's [Decisions API]() at 41.67% and [Jev]() 1.14 at 33.44%, run without web access or supplied external context. Uniform random guessing is 42.42%. LLMs with native search scored 96.56% to 97.33% at about 16.5 to 28.8 times the cost. Parallel generated these questions from its own customer topics; dataset terms prohibit redistribution. [source]() - **VENDOR-REPORTED**: The benchmark blog chart labels Baseten Elephant at 32.4% overall and 32.3% public. The repository records Baseten Mammoth at 45.0% overall and 37.0% public. [source]() #### Links - [Source: Announcement]() - [Source: Docs]() - [Source: Benchmark]() - [Source: Benchmark announcement]() ### Kev [Kev]() By Jared Palmer · Released: 2026-10-01 · OPEN [Kev]() is Jared Palmer's family of open, Jev-compatible decision models, from 0.8B to 27B parameters. A pointer head on a Qwen base scores each option, and the API matches TypeSafe's, so their SDK works against a local [Kev]() server. [Kev]() 1.0, released 1 Oct 2026, pins the family's checkpoints, model cards, evaluation suites and serving code; it introduces no newly trained weights. #### Facts - Sizes: 0.8B, 4B, 9B, 27B - Licence: Apache-2.0 - Base: Qwen3.5 (27B: Qwen3.8) - Release: [Kev]() 1.0, 1 Oct 2026 - Stars: 9,758 on GitHub (GitHub API, 2 Nov 2026) - Price: $0.042 per million input tokens and $0 per million output tokens for jaredpalmer/kev-4b on OpenRouter (VENDOR-REPORTED, 2 Nov 2026) · [Compare prices]() - Validated context: VENDOR-REPORTED: 65,536 tokens for Kev-27B; 8,192 for 0.8B, 4B and 9B, measured on CUAD contracts #### Claims and sources - **VENDOR-REPORTED**: [Kev]() 1.0's breadth-v1 held-out-dataset test scored a chance-corrected index of 23.3 for 0.8B, 38.0 for 4B, 41.0 for 9B and 52.3 for 27B, against [Jev]()'s 54.0, in the maker's fp32 evaluation at shipped temperatures. No [Kev]() trained on these datasets. [source]() - **VENDOR-REPORTED**: Kev-27B's 65,536-token validated context is the longest tested state whose CUAD accuracy stays within three percentage points of its 8k-token accuracy at the 95% lower bound. The smaller models already miss that tolerance at 16k. [source]() #### Links - [Source: GitHub]() - [Source: Kev 1.0 release]() - [Source: Hugging Face]() - [Source: Kev 4B on OpenRouter]() - [Source: GitHub API: stars]() ### DecisionTune 1.0 [DecisionTune 1.0]() By DecisionTune · Hugging Face repository upload: 2026-10-04 · OPEN A local decision model built on ModernBERT-large: give it a state, a question and options, and it returns a probability for each option in one encoder pass. It handles choice and yes/no questions without generating text, through PyTorch, MLX or ONNX. The open weights are under Apache-2.0. The date is when the weights repository went up on Hugging Face. #### Facts - Size: 395M parameters, ModernBERT-large plus a 4 KB scoring head - Licence: Apache-2.0 - Context: 8,192 tokens; longer sequences are refused #### Claims and sources - **VENDOR-REPORTED**: The maker reports 29.57 on Decision Index 0.2.1 (raw 46.80) from one complete run, with a second seed scoring 29.13 (raw 45.96). These are the maker's runs, not an independent reproduction. [source]() #### Links - [Source: Hugging Face]() ### Tev1-4B [Tev1-4B]() By Together AI · Released: 2026-09-23 · OPEN Together's Jev-inspired experiment: a Qwen3.5-4B fine-tune that reads a state, a question and 2 to 24 lettered options. Together says plainly it is not a non-autoregressive runtime like [Jev](): it keeps a standard next-token head and generates its answer as one letter. #### Facts - Base: Qwen3.5-4B - Licence: Code MIT; weights licence being finalised as of 2 Oct 2026 - Run it: Together serverless, or Ollama #### Claims and sources - **VENDOR-REPORTED**: Together reports 880/1,000 (88.0%) and 300/300 (100%) on evaluation sets used during development (VENDOR-REPORTED). [source]() - **INDEPENDENT**: In jev-rsi's independent baseline test, the author measured untuned Qwen3.5-4B using [Tev1]()'s decision protocol through local Ollama at 673/1,000 (67.3%) on test and 130/300 (43.3%) on policy\_transfer. [source]() #### Links - [Source: Hugging Face]() - [Source: GitHub]() - [Source: Announcement]() - [Source: Untuned 4B baseline test]() ### GLiNER2.5-Decide [GLiNER2.5-Decide]() By Fastino · Released: 2026-09-24 · OPEN A 340M-parameter encoder decision model from the team behind GLiNER, fine-tuned from GLiNER2. It scores the allowed answers to typed questions in one pass and is small enough to run on a CPU. #### Facts - Size: 340M parameters - Licence: Apache 2.0 - Latency: 167 ms p50 on a Xeon CPU, 38 ms on a V100 (Fastino-reported; batch 1, 64-token input) #### Claims and sources - **VENDOR-REPORTED**: Best average across 17 datasets on Fastino's own benchmark: 60.1% in the launch post, 60.2% on the model card as of 2 Oct 2026 [source]() #### Links - [Source: Announcement]() - [Source: Hugging Face]() ### Laya [Laya]() By Convai Innovations · Released: 2026-09-18 · OPEN A non-autoregressive decision engine on ModernBERT-large (421M), with a 322M multilingual sibling. Its author, Nandakishor Mukkunnoth, says he built the same approach a year before [Jev]() launched. #### Facts - Size: 421M (English), 322M (multilingual) - Licence: Apache 2.0 - Latency: 32.8 ms on a single GPU (maker-reported) #### Claims and sources - **VENDOR-REPORTED**: 6 to 8 times faster than [Jev]() [source]() - **VENDOR-REPORTED**: Built and open-sourced a non-autoregressive decision model a year before [Jev]() [source]() #### Links - [Source: Site]() - [Source: GitHub]() ### Nimble [Nimble]() By Bespoke Labs · Released: 2026-09-18 · OPEN A 9B decision model, LoRA-tuned from Qwen3.5-9B, with its 2,676 training examples published. One of the first two models in Ollama's decision support. Bespoke Labs followed with hosted [Nimble]() v4 (9B) and v4-max (27B) on 13 Oct 2026. #### Facts - Size: 9B - Licence: Apache 2.0 - Run it: ollama pull [nimble]() #### Claims and sources - **VENDOR-REPORTED**: Matches 90.1% of reference labels on 324 held-out examples, against 93.2% for [Jev]() 1.13 [source]() #### Links - [Source: GitHub]() - [Source: Ollama]() - [Source: Hugging Face]() - [Source: Nimble README]() ### Decision 1.0 [Decision 1.0]() By vLLM Semantic Router · Released: 2026-09-22 · OPEN Open decision models from the vLLM Semantic Router project, published as "[Decision 1.0](): towards open decision foundation models". Six were announced on 22 Sep. The follow-up, Decision 3.0 (d3), is Apache-2.0 and reads text, images and video, with sizes up to 27B. #### Facts - Sizes: 0.6B to 9B - Licence: Apache-2.0 (checked on Nox-4B) #### Claims and sources #### Links - [Source: Hugging Face]() - [Source: Announcement PR]() - [Source: d3 on Hugging Face]() ### Decider [Decider]() By Perplexity · Released: 2026-10-06 · OPEN Perplexity's decision model reads text and images and returns a probability for every answer. It ships as Apache-2.0 open weights and as a hosted API. V1.1 (27B) halved the price on 6 Oct; V1.2 and a 4B [Decider]() Flash followed on 20 Oct. #### Facts - Latest: [Decider]() V1.2 (27B) and [Decider]() Flash V1.2 (4B) - Price: $0.02 (V1.2) and $0.01 (Flash) per million input tokens on OpenRouter, output free (as of 22 Oct 2026) · [Compare prices]() - Licence: Apache-2.0 #### Claims and sources - **INDEPENDENT**: [Decider]() V1.1-27B was the most accurate of seven decision models (94.7%) on 799 unpublished yes/no questions about coding-agent work and destructive shell commands, and the only one to pass SILT's registered calibration rule. [Jev]()'s result was inconclusive. So were the results for OpenAI's [Decisions API]() and [Laya]() (11 Oct 2026) [source]() #### Links - [Source: V1.2 announcement]() - [Source: Hugging Face]() - [Source: V1.1 announcement]() ### Drex 2.0 [Drex 2.0]() By Nace.AI · Released: 2026-10-06 · OPEN [Drex 2.0]() returns probabilities over typed answers for text, with open weights under Nace.AI's Open RAIL licence. [Drex Max]() (13 Oct) adds images and screenshots; its 27B weights are now public under Nace.AI Open RAIL-M. [Drex Max Finance]() followed on 01 Nov 2026 as a hosted model. The entry's date is [Drex 2.0]()'s release in Nace.AI's model table. #### Facts - Size: [Drex 2.0](): 35B, about 3B active; [Drex Max](): dense 27B - Price: $0.06 per million input tokens for [Drex 2.0](); $0.03 for [Drex Max](), Max v1.1 and Max Finance, hosted · [Compare prices]() - Context: [Drex 2.0]() state up to 131,072 tokens; Max Finance whole request up to 262,144 tokens - Licence: [Drex Max]() weights: Nace.AI Open RAIL-M - Access: [Drex Max]() supports text and images through Python on NVIDIA GPUs. CPU and Apple Silicon runtimes support text only. Max Finance API id: drex-max-finance. #### Claims and sources - **VENDOR-REPORTED**: Nace.AI reports 85.7% accuracy across 3,361 decisions on 48 held-out finance benchmarks for [Drex Max Finance](), with a Finance Index of 77.3. These are its own 02 Nov 2026 launch figures. [source]() #### Links - [Source: Model table]() - [Source: Hugging Face]() - [Source: Drex Max weights and local inference]() - [Source: Drex Max Finance API guide]() ### H2O-Lightning-4B [H2O-Lightning-4B]() By H2O.ai · Released: 2026-10-07 · OPEN A 4B open model built on Qwen3.5-4B that runs on unmodified vLLM. Version 1.2.2 takes up to four images per request. #### Facts - Size: 4B - Licence: Apache-2.0 #### Claims and sources - **VENDOR-REPORTED**: Its text-only v1.1 ranked #1 of 104 open-weight systems on [JevBench]() v1.6.1's composite, 72.5 to [Jev]() 1.13's 71.5, read 7 Oct 2026 [source]() #### Links - [Source: Announcement]() - [Source: Hugging Face]() ### Jebadiah v2.1 [Jebadiah v2.1]() By frontier-infra · Released: 2026-10-10 · OPEN Open decision models that read option-label probabilities in a single pass with thinking switched off: you get a pick and a confidence, not an essay. GGUF and MLX builds sit next to the full weights. Released 10 Oct 2026, written up on 16 Oct. #### Facts - Sizes: 9B and 27B - Licence: Apache-2.0 #### Claims and sources #### Links - [Source: Release post]() ### Lumma-Fev-27B [Lumma-Fev-27B]() By FrontiersMind · Released: 2026-10-14 · OPEN A 27B open decision model that answers in English, Hindi, Tamil and Kannada. The makers say [Jev]() still leads on some reasoning benchmarks. #### Facts - Size: 27B, full fine-tune - Licence: Apache-2.0 #### Claims and sources - **VENDOR-REPORTED**: 90.7% vs [Jev]()'s 83.9% on the maker's held-out decision set [source]() #### Links - [Source: Release post]() ### Granite Switch 4.1 [Granite Switch 4.1]() By IBM · Released: 2026-10-15 · OPEN IBM's switchable models do chat, decisions and speech-to-text from one endpoint. 4.1 adds a general-purpose Decision adapter that speaks the /v1/systemone interface. #### Facts - Sizes: 8B and 30B - Licence: Apache-2.0 #### Claims and sources #### Links - [Source: Announcement]() ### TYGER [TYGER]() By NineNineSix · Released: 2026-10-17 · OPEN A decision model for speech. It listens to the microphone and your text-to-speech at the same time and keeps deciding whether to speak, keep listening, yield or drop in an "mm-hm", scoring audio in 80 ms frames streamed in 320 ms hops. Built for full-duplex voice agents, with an open-source agent (tyger-s2s) on top. #### Facts - Base: NVIDIA Nemotron 3.5 streaming ASR 0.6B - Licence: OpenMDW-1.1 - Works with: Any LLM and TTS (maker's claim) #### Claims and sources #### Links - [Source: Hugging Face]() - [Source: Announcement]() - [Source: Voice agent]() ### Talos [Talos]() By Tractian · Repository and weights upload: 2026-10-22 · OPEN A decision model that stands between AI agents and the controllers that run factories. Every Modbus or OPC UA request goes through one gate, where a small model decides in a single pass to allow it, answer from cache, defer it or deny it, and every decision is logged. The date is when the repo and weights went up. #### Facts - Size: 0.8B, built on Qwen3.5-0.8B - Licence: Apache-2.0 #### Claims and sources - **VENDOR-REPORTED**: About 5 ms per decision on a laptop CPU [source]() #### Links - [Source: GitHub]() - [Source: Hugging Face]() ### Mammoth 1 [Mammoth 1]() By Baseten (Base Labs) · Released: 2026-10-21 · OPEN Two open decision models from Base Labs, Baseten's research team: [mammoth-1]() (27B) and elephant-1 (9B), post-trained from Qwen3.5. They read text and images and return a probability for every option of each question, with all questions sharing one prefill of the state. Baseten launched them claiming the top spot on the community Jev Decision Index. On 29 Oct a third, smaller model joined: joey-1 (2B, built on Qwen3.5-2B), pitched as the fastest and cheapest. #### Facts - Sizes: [mammoth-1]() 27B (Qwen3.5-27B), elephant-1 9B (Qwen3.5-9B) and joey-1 2B (Qwen3.5-2B), 16,384-token context - Licence: MIT - Price: OpenRouter lists [mammoth-1]() at $0.10, elephant-1 at $0.035 and joey-1 at $0.01 per million input tokens, output free (as of 30 Oct 2026) · [Compare prices]() #### Claims and sources - **VENDOR-REPORTED**: Baseten says [mammoth-1]() ranked #1 of 208 models on Jev Decision Index 0.4 with a full score of 71.24, and elephant-1 is the best decision model under 15B parameters (62.65) [source]() - **INDEPENDENT**: In the index's current public data (Decision Index 0.4.1, generated 22 Oct 2026), [mammoth-1]() is first of 228 on both public-benchmark scores (75.34 skill, 81.19 raw) but third on the full score (68.44), behind [GLiDE]() v1.2 (71.52) and jeffan-31b (69.75) [source]() - **VENDOR-REPORTED**: Baseten says joey-1 has a P50 latency of 20 ms, 5x faster than comparable decision models [source]() #### Links - [Source: Announcement]() - [Source: Technical report]() - [Source: mammoth-1 on Hugging Face]() - [Source: elephant-1 on Hugging Face]() - [Source: joey-1]() ### Qev 0.8B [Qev 0.8B]() By labqoat · Hugging Face weights upload: 2026-10-20 · OPEN A tiny open decision model that runs locally, built on Qwen3.5-0.8B and reading text and images. The date is when the weights went up on Hugging Face. #### Facts - Size: 0.8B, built on Qwen3.5-0.8B-Base, about 1.3 GB of RAM on a Mac in 8-bit (maker's figure) - Licence: Apache-2.0 - Runs on: llama-server (GGUF), MLX on Apple silicon, or a System One API #### Claims and sources - **VENDOR-REPORTED**: First of 27 decision models of around 1B parameters on the public split of Benchmark Heaven (77.1%), and the fastest at a median 3.5 ms per request (batch size 1, one NVIDIA L40S), in the maker's own comparison [source]() #### Links - [Source: Model page]() - [Source: Hugging Face]() ### Decido [Decido]() By decido-ai · Weights upload: 2026-10-22 · OPEN An open decision model built to run on a laptop CPU: a LoRA fine-tune of Gemma 4 26B-A4B shipped as an 8-bit GGUF for llama.cpp, answering in the same wire format as [Jev]()'s System One API. The date is when the weights went up. #### Facts - Size: Gemma 4 26B-A4B (about 4B active parameters), 8-bit GGUF, about 30 GB of RAM - Licence: Apache-2.0 - Speed: A median 4.1 s for a 1,000-token document and 5 questions on an Apple M4 Pro CPU (maker's figure) - Code: The decido-ai/decido GitHub repository was gone on 31 Oct 2026; the weights and model card remain on Hugging Face #### Claims and sources - **VENDOR-REPORTED**: Scores a six-benchmark mean of 85.38 against [Jev]() 1.13's published 83.85, or 82.04 with the benchmarks' public training splits left out; the maker notes [Jev]()'s number is its published reference run, not [Jev]() run through the same harness [source]() #### Links - [Source: Hugging Face]() ### Arc-decide [Arc-decide]() By Arc53 (DocsGPT) · Hugging Face v0.2 upload: 2026-10-11 · OPEN The small decision model inside DocsGPT, used for quick judgements such as whether a message is a prompt injection or whether a chunk answers the question. It runs on CPU with ONNX Runtime. The date is when v0.2 went up on Hugging Face. #### Facts - Sizes: arc-decide-0.4b-v0.2 for CPU; arc-decide-0.6b-v0.1 for NVIDIA GPUs with vLLM - Licence: MIT, built on Qwen3 (Apache-2.0) #### Claims and sources #### Links - [Source: Docs]() - [Source: Hugging Face]() ### Valen [Valen]() By Valen Team · First Hugging Face weights upload: 2026-10-05 · OPEN Open decision models in three sizes that read text, images and video, built on Qwen3.5. They answer through both [Jev]()'s System One format and OpenAI's Decisions format, with a free web demo and API. The date is when the first weights went up on Hugging Face; the 4B followed on 6 October. #### Facts - Sizes: 0.8B, 2B and 4B, built on Qwen3.5, with ONNX and GGUF builds - Licence: Apache-2.0 - Inputs: Text, images and video #### Claims and sources #### Links - [Source: Site and demo]() - [Source: Hugging Face]() ### Fovea-4B [Fovea-4B]() By Sivasuthan · Weights upload: 2026-10-24 · OPEN An open decision model for images: give it a photo, document, receipt, chart or screenshot plus a typed question and it returns a probability for every allowed answer, served over [Jev]()'s System One API. The date is when the weights went up. #### Facts - Size: 4B, built on Qwen3.5-4B - Licence: Apache-2.0 #### Claims and sources - **VENDOR-REPORTED**: Scores 80.6 on the maker's own 2,360 held-out multiple-choice questions from public benchmarks, scored with ImageJevBench's formula, ahead of [H2O-Lightning-4B]() (79.9), Imajev-4B (79.4) and Jev-Vision 8B (75.7); the maker notes this is not the ImageJevBench leaderboard [source]() #### Links - [Source: GitHub]() - [Source: Hugging Face]() ### homi [homi]() By egma-ai · Hugging Face weights upload: 2026-10-27 · OPEN An open decision model that listens instead of reading: give it an audio clip and a typed question (how does the speaker feel? does the pitch rise?) and it returns a probability for every option, without a transcript. It is Qwen3-Omni-30B-A3B with frozen weights and a decision head in place of the text output. The date is when the weights went up on Hugging Face. #### Facts - Base: Qwen3-Omni-30B-A3B-Instruct, frozen, with a decision head and 17M parameters of LoRA - Licence: Apache-2.0 - Inputs: Audio clip plus Noul, Choice or Score questions #### Claims and sources - **VENDOR-REPORTED**: 73.9 on VocalGrad against 63.0 for Gemini 3.8 Flash, and 83.1 on MMSU against Gemini's 85.7. Cost per 10,000 MMSU questions: an estimated $0.28 of batched GPU time on one H200 at $4.54/hour, against $14.80 in Gemini API charges at promotional [pricing]() [source]() #### Links - [Source: GitHub]() - [Source: Hugging Face]() - [Source: Show HN]() ### Nimbus [Nimbus]() By Crele · Hugging Face weights upload: 2026-10-30 · OPEN A decision model for live audio. You ask yes/no questions in plain English (is music playing? is someone speaking?) and it answers every 40 ms of a stream with yes, no or can't tell, real time on a laptop CPU. The date is when the weights went up on Hugging Face. #### Facts - Size: About 97M parameters on the audio, plus a 0.6B text encoder (Qwen3-Embedding-0.6B) that reads each question once - Licence: Apache-2.0, code and weights - Inputs: A live audio stream plus yes/no questions; answers yes, no or cant\_tell per 40 ms #### Claims and sources - **VENDOR-REPORTED**: Says yes a median 0.24 s after a sound starts over the sounds it caught on 500 real AudioSet clips (speech 0.21 s, 97% of 181 speech onsets caught); reworded questions it never trained on scored 0.892 mean AUROC against 0.911 on its own wording [source]() #### Links - [Source: GitHub]() - [Source: Hugging Face]() ### GLiDE [GLiDE]() By Fastino · Weights repository upload: 2026-10-19 · OPEN Fastino's big decision model, built on Qwen3.8-27B: give it text, JSON or images plus typed questions and it returns a probability for every option, with nothing generated. The weights are on Hugging Face under Fastino's own community licence, and Fastino also hosts it as an API. The date is when the weights repository went up; the version scored on the index is 1.2. #### Facts - Base: Qwen3.8-27B; 52.2 GB of bf16 weights, needs one 80 GB GPU - Licence: Fastino community licence (not a standard open-source licence) - Access: Weights on Hugging Face, or Fastino's hosted System One API #### Claims and sources - **VENDOR-REPORTED**: Fastino says [GLiDE]() 1.2 is #1 of 229 models on Decision Index 0.4.1 (as of 22 Oct 2026) [source]() - **INDEPENDENT**: In the index's public data (release 0.4.1, generated 22 Oct 2026), [GLiDE]() 1.2 is first of 228 entries on the full score (71.52), ahead of jeffan-31b (69.75) and [mammoth-1]() (68.44); [mammoth-1]() leads both public-benchmark scores. In the index's 29 Oct data, [GLiDE]() is second of 247, behind [KnowLine]() 31B Gen1 (71.86) [source]() #### Links - [Source: Hugging Face]() - [Source: API docs]() - [Source: Decision Index]() ### UltraBERT Curation [UltraBERT Curation]() By Daniel van Strien · Released: 2026-10-28 · OPEN A 380M-parameter decision model for data curation: give it a web page, prompt, maths problem or library record and typed questions (topic, format, educational value, difficulty) and it returns a probability for every option. The release centres on its recipe: the data-building and training scripts are public, with source and labeller/licence metadata where known, and 516,964 of the 535,730 training rows are released (web text omitted). #### Facts - Size: 380M parameters, UltraBERT-380M encoder plus a GLiClass scoring head - Licence: Apache-2.0 - Open data: 516,964 of 535,730 training rows released, plus 1,020,000 FineWeb pages scored with it #### Claims and sources #### Links - [Source: Hugging Face]() - [Source: Recipe]() ### KnowLine [KnowLine]() By PelaAI · First Hugging Face family release: 2026-10-07 · OPEN A fast-moving family of open decision models from PelaAI, trained in an agent-driven data loop, with English and Chinese in mind: from 0.8B up to 35B-A3B, with new generations posted every few days. The newest big one, KnowLine-31B-Gen1 on a Gemma 4 base, went up on 23 Oct. The date is the first family release on Hugging Face (4B-Gen1). #### Facts - Sizes: 4B through Gen5.5; 26B-A4B through Gen2.5; 35B-A3B through Gen1.5; 0.8B, 2B, 9B and 31B (Gemma 4 base) at Gen1; community GGUF builds - Licence: Apache-2.0 - Languages: English and Chinese #### Claims and sources - **INDEPENDENT**: In the Decision Index's 29 Oct 2026 data, [KnowLine]() 31B Gen1 is first of 247 on the full score (71.86), ahead of [GLiDE]() 1.2 (71.52) [source]() - **VENDOR-REPORTED**: PelaAI reports 0.4% of answers changed by an instruction planted in the state, 5.3% with injection templates unseen in training [source]() #### Links - [Source: Hugging Face]() - [Source: Decision Index]() ### Ollama v0.35 [Ollama v0.35]() By Ollama · Released: 2026-09-28 · RUNTIME Ollama added native decision models through /v1/systemone, based on TypeSafe's [Jev]() API, so you can run one on your own machine. It shipped with [Nimble]() and [Tev1](). #### Facts - Endpoint: /v1/systemone - Question types: choice, noul, score - Launch models: [Nimble](), [Tev1]() #### Claims and sources #### Links - [Source: Release notes]() ### SemIf (formerly OpenJev) [SemIf (formerly OpenJev)]() By TheoLeeCJ · COMMUNITY Reproduces [Jev]()'s interface pattern with open models: it reads typed option probabilities directly from a model, with no answer sentence and no decoding loop. Runs in your browser, or at home on a GPU, Apple Silicon or CPU. Independent of TypeSafe, and it does not reproduce [Jev]()'s model or training. #### Facts - Availability: Runs in the browser - Models: Qwen3 0.6B, MiniCPM5 2B or Qwen3.5 4B, as pinned GGUF builds through wllama #### Claims and sources - **CODE-VERIFIED**: Checked against its own README on 2 Oct 2026 [source]() - **INDEPENDENT**: On [JevBench]() v1.8.6, published 01 Nov 2026, [SemIf]() (formerly [OpenJev](), Qwen3.5-4B) ranks #143 at 55.4 Capability, with 27.1 Intelligence and 83.6 Calibration. [Jev]() 1.13.0 is the unranked reference at 77.1 Capability. [source]() #### Links - [Source: openjev.com]() - [Source: TheoLeeCJ/SemIf-OpenJev]() ### bnsd55/jevmlx [bnsd55/jevmlx]() By bnsd55 · COMMUNITY Parallel constrained decisions for local models through mlx-lm on Apple Silicon: the schema is a set of typed fields, the context is prefilled once and broadcast across one batch row per field. MIT licensed. Large answer sets may be chunked rather than done in a single pass. #### Facts - Availability: Open source #### Claims and sources - **CODE-VERIFIED**: Checked against its own code and README on 18 Sep 2026 [source]() #### Links - [Source: GitHub]() ### kw2828/OpenJev [kw2828/OpenJev]() By kw2828 · COMMUNITY A browser and local playground for decision experiments on memory, uncertainty and control, using Qwen-family models and returning option probabilities. MIT licensed. Its README states it is independent of [Jev](), of [openjev]().com and of zhihz/openjev. #### Facts - Availability: Open source #### Claims and sources - **CODE-VERIFIED**: Checked against its own code and README on 18 Sep 2026 [source]() #### Links - [Source: GitHub]() ### akash-kamat/system-one-gemma [akash-kamat/system-one-gemma]() By akash-kamat · COMMUNITY Gemma 3 270M with a scoring head, producing a decision in a single forward pass with no text generation. Code and adapter weights are publicly readable, but the repository carries no licence, so it is not open source in the legal sense. The Gemma base weights remain under their own separate terms. #### Facts - Availability: Source available; no licence found as of 18 Sep 2026 - Model: 270M (Gemma 3 backbone) #### Claims and sources - **CODE-VERIFIED**: Checked against its own code and README on 18 Sep 2026 [source]() #### Links - [Source: GitHub]() ### leofmarciano/groot [leofmarciano/groot]() By leofmarciano · COMMUNITY Groot 1 S, a 149M-parameter open typed-decision model that speaks the System One request format and reads states of up to 64k tokens in the maker's tests. Its forward pass also ships in the Bend language. #### Facts - Size: 149.3M - Licence: Apache-2.0 #### Claims and sources - **VENDOR-REPORTED**: 75.2% vs [Jev]()'s 73.7% on the full 2,000-question test split of the public LocalLLaMA/typed-decisions set, run by the maker, with a Brier score under half of [Jev]()'s [source]() #### Links - [Source: GitHub]() ### jwuthri-cloud/SelfJev [jwuthri-cloud/SelfJev]() By jwuthri-cloud · COMMUNITY A self-hosted decision model with a Jev-compatible API: you run the server on your own GPU, and a 230 MB vision adapter on Qwen3.5-4B reads text and images. Python SDK, server and fine-tuning in one package. #### Facts - Base: Qwen3.5-4B - Licence: Apache-2.0 (code) #### Claims and sources #### Links - [Source: GitHub]() - [Source: Hugging Face]() ### prismhq/jev-router [prismhq/jev-router]() By prismhq · TOOLING A LiteLLM-based router with an optional [Jev]() decision backend: when a TypeSafe API key is configured it can use [Jev]() to pick a model per request, and otherwise falls back to its own rules-based [decider](). MIT licensed. #### Facts - Availability: Open source #### Claims and sources - **CODE-VERIFIED**: Checked against its own code and README on 18 Sep 2026 [source]() #### Links - [Source: GitHub]() ### mizorewww/laya-mlx [mizorewww/laya-mlx]() By mizorewww · TOOLING Runs the open [Laya]() decision models natively on Apple Silicon with MLX: no PyTorch, no cloud API, zero output tokens. Its demo plays Snake, calling [Laya]() for every move with a safety layer that can override unsafe ones. Apache-2.0, and more starred than any project on our list except browser-use's jev-ultrafast. #### Facts - Stars: 6,994 on GitHub (GitHub API, 22 Oct 2026) - Licence: Apache-2.0 #### Claims and sources - **VENDOR-REPORTED**: 13.4 ms median for one short English typed decision, 7.4 ms with the multilingual checkpoint, on the maintainer's M3 Max [source]() #### Links - [Source: GitHub]() ### Hume Audio Decisions API [Hume Audio Decisions API]() By Hume AI · Released: 2026-11-02 · HOSTED Hume's hosted model answers questions about speech and sound with probabilities. Send one MP3 or WAV recording, 5 seconds to 5 minutes long, with questions in plain language. It supports yes/no, choice, score and time-span answers, and returns no transcript. Hume says the probabilities estimate how human raters would respond. #### Facts - Access: Playground; request API access through Hume's waitlist - Inputs: One MP3 or WAV per request, 5 seconds to 5 minutes; the model hears 16 kHz mono audio - Price: $0.47 per million input tokens (audio, questions and context); answers not billed. Hume's billing docs, checked 03 Nov 2026 · [Compare prices]() - SDKs: Python hume-audio-decisions and TypeScript @humeai/audio-decisions; Python SDK code MIT #### Claims and sources - **VENDOR-REPORTED**: Hume reports 70.7% agreement with a human panel's average judgment, against 56.5% for an individual-rater baseline, in its initial evaluation. [source]() - **VENDOR-REPORTED**: On Hume's product-page chart, its model averages 0.82 across 10 benchmarks at 0.23 seconds median latency per screen. The chart gives Gemini 3.8 Flash 78%, Qwen3.8-Omni-Flash 76% and Inkling-Small 67%; it names MMAU, ParaPair and EmphAssess among the benchmarks. [source]() #### Links - [Source: Announcement]() - [Source: Product and playground]() - [Source: Docs]() - [Source: Billing]() - [Source: Audio preparation]() - [Source: Launch changelog]() - [Source: Python SDK]() ### Zoom Decisions API [Zoom Decisions API]() By Zoom · Released: 2026-10-19 · HOSTED [Zoom]()'s hosted [Decisions API]() and [Decisions API Pro]() accept text, images or both. Define several questions about the same input and get structured answers in one call for condition checks, option selection and scoring. #### Facts - Models: [Decisions API]() and [Decisions API Pro]() - Price: $0.042 per million tokens for [Decisions API](); $0.10 for Pro, per [Zoom]()'s product page checked 03 Nov 2026 · [Compare prices]() - Endpoint: POST /v2/aiservices/decisions/decide - Access: [Zoom]() AI Services portal and playground; sign-up or login required #### Claims and sources - **VENDOR-REPORTED**: [Zoom]() reports 83.9% overall accuracy for [Decisions API Pro]() on a comparison it says uses [JevBench]() and ImajevBench. Its product page says Pro leads the tested models but gives no dated per-model results table. [source]() #### Links - [Source: Product, playground and pricing]() - [Source: Launch changelog]() ### LingDT-Decision-1.0 [LingDT-Decision-1.0]() By Ant Digital Technologies (蚂蚁数科) · Released: 2026-10-22 · HOSTED Ant Digital Technologies' System One model for choosing answers to business questions, including native multi-select and decisions ordered by dependencies. IT之家 reported its 22 Oct 2026 launch in 4B and 27B versions for customer service, risk and finance. Ant's current DTMaaS listing describes the [lingdt-decision-1.0]() endpoint as a 4B test model. #### Facts - Sizes: 4B and 27B announced, per IT之家; the current DTMaaS endpoint identifies itself as 4B - Try it: Open the DTMaaS model page and its API guide. Use an API key with POST maas-api.antdigital.com/v1/systemone and model [lingdt-decision-1.0](); send state, questions and optional DAG dependencies - Limits: Ant lists 32 questions per request, 255 options, DAG depth 3 and a default 60 requests per minute. Its English and Chinese descriptions disagree on whether the 16K context limit applies per question or per request. - Price: Ant lists ¥0 per million input and output tokens, with an original input rate of ¥0.20 per million tokens. Checked 04 Nov 2026. The rate is for the 4B test endpoint during a free trial; no trial end date was established. · [Compare prices]() - Compare: [Jev]() and [Microsoft-Decision-1]() are hosted alternatives, each listed at $0.042 per million input tokens. LingDT's free-trial price covers a 4B test endpoint; these rates do not compare accuracy #### Claims and sources - **VENDOR-REPORTED**: Across 23 open benchmarks, Ant says [LingDT-Decision-1.0]() beats [Microsoft-Decision-1]() and [Jev]() 1.13, with the 27B version most accurate in its size class. These are Ant's results as reported by IT之家, not an independent comparison. [source]() - **VENDOR-REPORTED**: IT之家 reports Ant's claim of over 90% fewer tokens and up to 10x faster responses than LLMs in its customer-service, risk and finance tests. [source]() #### Links - [Source: DTMaaS model page]() - [Source: Official model data and API example]() - [Source: API guide]() - [Source: Launch report (IT之家)]() - [Source: Jev pricing source]() - [Source: Microsoft-Decision-1 source]() ### Ru-Jev [Ru-Jev]() By besktech (beskvladimir-create) · First Hugging Face repository creation: 2026-11-01 · OPEN A Russian-first typed-decision model built on mmBERT-base (307M), with a separate scoring head. It runs on CPU and supports yes/no, choice and ordinal scales. The maker says it is unaffiliated with TypeSafe AI and was not trained on [Jev]() outputs. Hugging Face now documents v1.1; the GitHub README still reports v1 results. #### Facts - Size: 307M (mmBERT-base) - Licence: Code MIT; weights tagged MIT on Hugging Face - Run it: Load the backbone, head.pt and calibration.json using the Hugging Face card's Python example; the standard inference widget cannot load the separate head - Release date: First Hugging Face repository creation: 01 Nov 2026; v1.1 card checked 03 Nov 2026 #### Claims and sources - **VENDOR-REPORTED**: For v1.1, the maker reports 0.825 micro accuracy and 0.0200 ECE on 13,373 decisions, half the validation split, with temperature fitted on the other half. Its 0.0159 ECE fitted and scored on the same data is in-sample. [source]() - **VENDOR-REPORTED**: The v1.1 card reports full-validation accuracy of 0.965 on 60-intent routing and 0.954 on 18-scenario routing, but 0.522 on reasoning questions and 0.450 on three-way NLI. Category wording also changes routing accuracy. [source]() #### Links - [Source: Weights and v1.1 model card]() - [Source: GitHub]() - [Source: Release metadata]() ### Intern-Decision [Intern-Decision]() By InternLM · First Hugging Face upload: 2026-09-26 · OPEN Send a shared state, named questions and optional images to these multimodal models. A single forward pass returns an answer distribution for every question, without generating text. The Qwen3.5-based family has 0.8B, 2B and 4B sizes. #### Facts - Sizes: 0.8B, 2B and 4B - Base: Qwen3.5; the 4B model uses Qwen3.5-4B - Licence: Apache-2.0 - Inputs: Text and images #### Claims and sources - **VENDOR-REPORTED**: InternLM reports a seven-benchmark average of 90.02 for [Intern-Decision-4B]() against 88.74 for [Jev](), with Brier / ECE of 0.347 / 0.065 against 0.358 / 0.095. [source]() - **VENDOR-REPORTED**: The model card reports mean latency of 44.16 ms for the 4B on one RTX 4090, against 109.70 ms for [Jev](). The local and network paths differ, so these timings are not like for like and depend on hardware and workload. [source]() #### Links - [Source: 4B weights and model card]() - [Source: GitHub]() - [Source: Report]() - [Source: Hugging Face upload metadata]() ### MiniCPM-OCR Jev [MiniCPM-OCR Jev]() By OpenBMB · First Hugging Face upload: 2026-10-23 · OPEN This 2B specialist scores whether a candidate OCR reading matches a document region. It handles text, formulas and tables, returning a score from 0 to 1. Qwen3.5-2B supplies the backbone; a binary classification head replaces its language head. Scores can help select among OCR engines or route uncertain regions to a person. #### Facts - Size: 2B - Base: Qwen3.5-2B - Licence: Apache-2.0 - Access: Open weights in BF16, AWQ, GPTQ and GGUF; a free API on ModelBest's platform #### Claims and sources - **VENDOR-REPORTED**: OpenBMB reports decision AUROC ×100 of 93.6 on its JevOCRBench, against 90.1 for a row named JEV-35B-A3B and 93.8 for Qwen3.5-397B-A17B. The card does not identify JEV-35B-A3B further. [source]() - **VENDOR-REPORTED**: Across four OCR parsers on the same 200 samples with manually labelled readings, OpenBMB reports AUROC ×100 from 91.9 to 93.4. These are its own parser and sample combinations. [source]() #### Links - [Source: Weights and model card]() - [Source: Code and deployment]() - [Source: JevOCRBench]() - [Source: Hugging Face upload metadata]() ### hs-decide-1-flash [hs-decide-1-flash]() By Vectorize (Hindsight) · Released: 2026-11-03 · OPEN Vectorize's Hindsight team fine-tuned this 0.8B model from Qwen3.5-0.8B for agent memory. It returns probabilities for choice, score and yes/no questions in one pass. Hindsight first tried these judgments with [Jev](), then trained its own model to check whether facts are new and evidence is enough. It uses the same System One request shape. #### Facts - Size: 0.8B - Base: Qwen/Qwen3.5-0.8B - Licence: Apache-2.0 - Inputs: Text and structured state - Languages: English, Spanish, French, German, Portuguese, Japanese, Korean and Chinese - Context: 65,536 tokens per request (trained) - Price: $0.02 per million input tokens on Hindsight Inference; input only · [Compare prices]() - Access: Open weights; vLLM with the repository's serving plugin; hosted POST /v1/systemone #### Claims and sources - **VENDOR-REPORTED**: In Hindsight's own 1.2.0 evaluation, quality rises for six of seven LLMs. glm-5.3-flash goes from 73.8 to 87.1 and deepseek-v4.1-flash from 81.9 to 85.1; gemini-3.1-flash-lite goes from 87.1 to 86.6. That model is also the judge, so its rows are self-graded. Costs fall for five LLMs. [source]() - **VENDOR-REPORTED**: glm-5.3-flash with the decision model scores 87.1 at $7.72 per thousand cases, against 89.0 at $147 for gemini-3.8-flash alone. Costs use list prices and an equal mix of four memory operations. [source]() - **VENDOR-REPORTED**: Median decision-model latency is 167 ms, with 350 ms at p90. The replaced LLM calls had median latencies of 1 to 7 seconds, depending on the model. [source]() - **VENDOR-REPORTED**: Reflect costs 4.6x less with gemini-3.8-flash, while quality drops from 94.1 to 91.2. With gemini-3.1-flash-lite, reflect quality drops from 94.1 to 85.3. [source]() - **VENDOR-REPORTED**: Against three other 0.8B models, its mean across eight benchmarks is 75.6, ahead of 74.5, 73.4 and 66.6. On the public Decision Index it trails StartLux-Decision 0.8B Instant, 36.1 to 53.8. [source]() #### Links - [Source: Launch and evaluation]() - [Source: Weights and model card]() - [Source: Hindsight code]() - [Source: System evaluation]() ### Smaug-Decisions [Smaug-Decisions]() By Abacus.AI · Maker blog announcement: 2026-11-02 · OPEN Abacus.AI's 4B decision model answers a typed question about audio, an image or text with a probability per option in one pass. It is a Gemma 4 E4B-it fine-tune. The family includes a 744M audio model and a LiteRT build for phones. Each 4B request takes one audio clip up to 30 seconds or one image, with up to 20 options. #### Facts - Size: 4B effective, 8.6B stored; 744M audio tier - Base: google/gemma-4-E4B-it; google/embeddinggemma-2 for the audio tier - Licence: Apache-2.0 - Inputs: Audio, images and text; 2 to 20 options - Formats: Transformers; LiteRT-LM; Core ML int8 for the audio tier #### Claims and sources - **VENDOR-REPORTED**: On Abacus.AI's 6,671-item Earshot benchmark, the 4B scores 60.1 audio skill with ECE 0.069. Gemini 4 Argon scores 59.4 with ECE 0.123, and Gemini 3.8 Flash 50.5 with ECE 0.204. The 0.7B audio tier scores 53.0 with ECE 0.086. [source]() - **VENDOR-REPORTED**: Using the 0.7B audio tier first and escalating below 0.9 matches the 4B's 60.1 audio skill at 130 ms mean latency instead of 210 ms in Abacus.AI's run. [source]() - **VENDOR-REPORTED**: Abacus.AI reports 57.9 on public Decision Index images and 56.6 on text. These are its own runs, not leaderboard submissions. [source]() - **VENDOR-REPORTED**: The 3.45 GB LiteRT build scores 59.4 audio skill in a PyTorch replica of its quantized weights. Abacus.AI measures about 0.6 seconds per text decision on an iPhone 14 Pro Max. [source]() #### Links - [Source: Weights and model card]() - [Source: 0.7B audio tier]() - [Source: LiteRT build]() - [Source: Maker blog announcement]() - [Source: X announcement: 3 Nov 2026]() ### REAX 1.3 [REAX 1.3]() By REAX · Released: 2026-11-03 · OPEN [REAX 1.3]() runs two model parts on one GPU and averages their per-option probabilities 50/50. One blends a Bittensor subnet 92 miner fine-tune with [KnowLine](); the other blends a miner fine-tune from [Mammoth-1]()'s lineage with [Decision 1.0]()'s d3-ultra. Its interface accepts Jev-style requests. The hosted service at decisionmodels.io and the Hugging Face account decisionmodels belong to a separate business unrelated to decisionmodels.ai. #### Facts - Licence: Apache-2.0 for the blended work, subject to preserved upstream licences and notices; REAX Engine is MIT - Weights: About 115 GB in FP8; reference hardware is one RTX PRO 6000 Blackwell 96 GB - Inputs: Text; noul, choice and score questions - Access: Open Hugging Face weights and REAX Engine reax13 mode - Local input cap: 32,000 proxy tokens, calculated from compact request JSON characters divided by four, rounded down; larger requests are refused. #### Claims and sources - **VENDOR-REPORTED**: REAX reports 73.71 on all 140,178 scoreable public Decision Index 0.3 requests. Counting the one request above the released engine's input cap as refused gives 73.70. These are internal measurements, not official leaderboard scores. [source]() - **VENDOR-REPORTED**: On 3,478 public items answered by all eight models, REAX reports 69.5 against 68.8 for d3-ultra, within noise, and 67.4 for [KnowLine](). The second draw was added after a partial result was known. The public subset cannot establish an official Full-score ranking. [source]() - **VENDOR-REPORTED**: On a uniform 1,000-call sample, REAX measures 207 ms median and 655 ms mean latency, one request at a time on one RTX PRO 6000. Long multi-question requests can take minutes. [source]() #### Links - [Source: Weights, results and upstream notices]() - [Source: REAX Engine]() - [Source: Announcement]() ### K-Decision-9B [K-Decision-9B]() By ONYULTEA · First Hugging Face upload: 2026-11-03 · OPEN A Korean-specialised 9B decision model based on Qwen3.5-9B, trained with int8 LoRA at rank 32. It reads a state and typed questions, then returns probabilities over the options in one forward pass. The first public release supports choice and yes/no questions through a Jev-compatible /v1/systemone server. #### Facts - Size: 9B - Base: Qwen/Qwen3.5-9B - Licence: Apache-2.0, declared for the model weights - Inputs: Text; choice questions with 2 to 26 options and yes/no questions - Language: Korean-specialised; the model card also tags English - Limits: State up to 6,000 characters; unsupported requests return HTTP 422 - Access: Public model.safetensors weights and serving code #### Claims and sources - **VENDOR-REPORTED**: ONYULTEA reports 0.866 accuracy on a five-task KoBEST conversion with 2,600 items, and 0.635 on 3,000 korean-jev-decision-index items. Its LocalLLaMA typed-decisions score of 0.818 is labelled fitted. [source]() - **VENDOR-REPORTED**: On the public Decision Index 0.3 suite, ONYULTEA reports index 30.7, raw 44.8 and 168 ms p50 latency. It says training and serving used one 24 GB consumer GPU. [source]() #### Links - [Source: Weights, licence declaration and results]() - [Source: Serving and evaluation code]() - [Source: Weight file inventory]() ## Public projects built on decision models Each one opened and checked against its own repo, with star counts as of the day each was added. [Jev](https://decisionmodels.ai/models/jev), released on 15 September 2026, is behind 82 of the 123 projects; the other 41 run on twenty-five different models, or models you load locally. Checked 03 Nov 2026, GitHub search returned [11,206 public repositories matching "jev"](https://github.com/search?q=jev+created%3A2026-09-15..2026-09-30&type=repositories) created from 15 to 30 September, and [27](https://github.com/search?q=jev+created%3A2026-09-11..2026-09-14&type=repositories) created from 11 to 14 September. ### Hindsight agent memory [Hindsight agent memory]() [Source: vectorize-io/hindsight]() Hindsight uses [hs-decide-1-flash]() to prune and deduplicate recall, check whether evidence answers a request, consolidate new facts and decide whether a memory could change a knowledge page. It first tried this with [Jev](). In the team's own Hindsight 1.2.0 evaluation, glm-5.3-flash plus the decision model scores 87.1 at $7.72 per thousand cases, against 89.0 at $147 for gemini-3.8-flash alone. Reflect quality falls with some LLMs. gemini-3.1-flash-lite judges all runs and also appears in the comparison. - Creator: Vectorize - Stars: 58,477 - Licence: MIT - Model: [hs-decide-1-flash]() - Tags: memory, agents Authors' published runs · VENDOR-REPORTED; not independently reproduced - [Source: Launch and evaluation]() - [Source: MIT licence]() ### Ultra-fast browser agent [Ultra-fast browser agent]() [Source: browser-use/jev-ultrafast]() Zurich to London on Google Flights in 7.1 seconds, including model calls, generated text, browser work and loading waits. Swapping to a decision model took median task time from 9.450s to 7.092s and median browser protocol calls from 1,092 to 101. One natural-language goal, on the real web, at 1x speed. Maker-reported: three repeats of one task on one browser profile. - median task, down from 9.45 s: 7.09 s - browser calls, down from 1,092: 101 - Creator: browser-use - Stars: 23,396 - Licence: MIT - Model: [Jev]() - Tags: browser, agent Figures from the project's own README and docs/performance.md. The maintainers note this is three repeats of one task on one browser profile, not a general benchmark. - [Source: GitHub]() - [Source: Measurements]() ### One trade decision per block [One trade decision per block]() [Source: jarrodwatts/jev-trader]() Attempts one trade decision per Monad block on the Kuru MON-USDC market, and holds when the model misses the block deadline. - Creator: Jarrod Watts - Stars: 3,050 - Licence: MIT - Model: [Jev]() - Tags: trading, real-time ### Racing game steered by d1 [Racing game steered by d1]() [Source: Liquid4All/cookbook]() A pixel-art racer where [d1]() picks left, centre or right on every tick, with a mode that races it against [Jev](). Liquid AI's own example, in their cookbook repo, so the star count is for the whole repo. - Creator: Liquid AI - Stars: 2,623 - Licence: no licence - Model: [d1]() - Tags: game, demo ### Computer use at $0.0002 a step [Computer use at $0.0002 a step]() [Source: awlevin/typesafe-computer-use]() OCRs the screen, then classifies the next action. The claimed per-step cost is the maintainer's own figure, not independently measured. - Creator: awlevin - Stars: 1,251 - Licence: MIT - Model: [Jev]() - Tags: computer-use, cost ### Staged code review [Staged code review]() [Source: devagrawal09/jev-review]() A staged review workflow with a local dashboard: each stage is a bounded judgement rather than a written critique. - Creator: Dev Agrawal - Stars: 686 - Licence: MIT - Model: [Jev]() - Tags: code-review, workflow ### Mario, played from emulator state [Mario, played from emulator state]() [Source: fhshaik/typesafe-mario]() An agent that plays Super Mario Bros. by classifying the next action from structured emulator state rather than generating a plan. - Creator: fhshaik - Stars: 456 - Licence: no licence - Model: [Jev]() - Tags: game, control ### Session recaps sorted by Solar Decide [Session recaps sorted by Solar Decide]() [Source: Sidiora-Labs/codify]() A local agent workflow tool written in C. Its recap command sends each session transcript to [Solar Decide](), which labels every statement as a goal, decision, constraint, done or noise before a chat model writes the summary. - Creator: Sidiora Labs - Stars: 433 - Licence: MIT - Model: [Solar Decide]() - Tags: agent, tooling ### Page clutter removal [Page clutter removal]() [Source: kitze/unclutter]() A browser extension that decides, element by element, what on a page is clutter, with reusable template rules. - Creator: Kitze - Stars: 382 - Licence: MIT - Model: [Jev]() - Tags: browser, extension ### Camera-input drone in simulation [Camera-input drone in simulation]() [Source: RomanSlack/jev-drone]() A drone experiment in MuJoCo with a small judgement model in the control loop, on camera input. The maintainer calls the harder tunnel run a partial result. - Creator: Roman Slack - Stars: 282 - Licence: MIT - Model: [Jev]() - Tags: robotics, control ### fastbrowse: a browser agent where Jev picks each action [fastbrowse: a browser agent where Jev picks each action]() [Source: agent-labs-dev/fastbrowse]() An LLM plans, reads and writes; [Jev]() picks each action from the controls indexed on the page, and code stops irreversible clicks without authorization. On 46 tasks run three times each against the hosted Browser Use agent, it passed 99.2% of runs to 96.3% and cost $0.0296 a run to $0.3180, 10.7x less, but was slower: 18.9 s a task to 15.3 s (published campaign release-0.7.1). - Creator: agent-labs-dev - Stars: 155 - Licence: MIT - Model: [Jev]() - Tags: browser, agent ### Live meter over video [Live meter over video]() [Source: ChetasLua/jevmeter]() Scores every sentence of a video and renders the running result as an overlay. - Creator: ChetasLua - Stars: 108 - Licence: MIT - Model: [Jev]() - Tags: video, scoring ### Paper-trading command center [Paper-trading command center]() [Source: mars-tw/jev-trading-command-center]() A paper-trading research app where [Jev]() makes the decisions, wrapped in an animated 3D command center with model pets and a ten-timeframe research panel. Every trade is simulated; there is no real order path. - Creator: mars-tw - Stars: 63 - Licence: MIT - Model: [Jev]() - Tags: trading, research ### Visual inspection workbench with Laya [Visual inspection workbench with Laya]() [Source: uu889/laya-opencv]() A local inspection workbench: OpenCV measures the image, hard rules decide anything numeric, and a [Laya]() decision model cross-checks the result. Built for fruit picking, weed and pest monitoring, and steel pipe and textile checks. Version 3.0 also trains your own Laya-format decision models. - Creator: uu889 - Stars: 58 - Licence: MIT - Model: [Laya]() - Tags: inspection, vision ### Full-duplex voice agent [Full-duplex voice agent]() [Source: nineninesix-ai/tyger-s2s]() A browser voice agent where the [TYGER]() decision model is the ears: it transcribes you and decides when you've finished, when you've cut in, and when to say "mm-hm", while an LLM writes the replies. - Creator: NineNineSix - Stars: 53 - Licence: Apache-2.0 - Model: [TYGER]() - Tags: voice, agent ### Frame-by-frame video QA [Frame-by-frame video QA]() [Source: levantolabs/sagecut-release]() QA for Remotion videos: code measures every frame, the Sage decision model answers yes/no questions about each one, and only the frames it flags go on to the slower critics. - Creator: Levanto Labs - Stars: 47 - Licence: MIT - Model: Sage - Tags: video, qa ### Commit classification [Commit classification]() [Source: devanshbatham/commit-miner]() Classifies git diffs and messages into bug fixes, security fixes with CWEs, and change types. - Creator: Devansh Batham - Stars: 39 - Licence: no licence - Model: [Jev]() - Tags: classification, security ### laya-guard: review shell commands before they run [laya-guard: review shell commands before they run]() [Source: scs0209/laya-guard]() Checks coding-agent tool calls with rules and a local multilingual [Laya]() fine-tune, returning allow, review or block. VENDOR-REPORTED, base to fine-tune: the model alone held 32 then 39 of 43 attacks and allowed 25 then 44 of 50 everyday commands. Six end-to-end verdicts improved and two worsened, missing a preset no-regression criterion. Accuracy on 162 held-out training rows rose from 0.611 to 0.852. The author inspected the benchmark and evaluation sets between training runs, so they are less independent than a fresh test. - Creator: scs0209 - Stars: 35 - Licence: MIT - Model: [Laya]() - Tags: guardrails, shell, local - [Source: Fine-tune and evaluation]() ### Vercel form router: send submissions to the right team [Vercel form router: send submissions to the right team]() [Source: vercel-labs/jev-ai-sdk-form-router]() Routes lead, contact and issue forms with [Jev]() through Vercel AI Gateway and the AI SDK. It accepts [Jev]()'s destination when confidence is at least 95%; otherwise, or if [Jev]() fails, openai/gpt-6-luna-fast evaluates the submission independently. The result shows the destination, deciding model, timings and an email preview, with optional delivery. - Creator: vercel-labs - Stars: 15 - Licence: MIT - Model: [Jev]() - Tags: forms, routing ### Decision models vs prediction markets [Decision models vs prediction markets]() [Source: yorkeccak/model-vs-market]() A live site where OpenAI Decisions, [Jev]() and Cloudflare's [Clef]() put a probability on live questions after reading recent news, never the odds. Their probabilities are compared against Polymarket and Kalshi prices. - Creator: yorkeccak - Stars: 12 - Licence: MIT - Model: [Decisions API](), [Jev](), [Clef]() - Tags: forecasting, benchmark ### Nouli: multi-turn sessions for decision models [Nouli: multi-turn sessions for decision models]() [Source: nouli-ai/nouli]() Decision models forget everything after each answer; Nouli keeps a session. Type facts as you learn them and start a line with ? to ask a new question; earlier facts and answers stay in the history. A Rust CLI and library that can chain sessions into decision graphs, with [Jev]() through its TypeSafe adapter or other providers. - Creator: nouli-ai - Stars: 12 - Licence: MIT - Model: [Jev]() - Tags: cli, rust ### Guardrails for coding agents' bash commands [Guardrails for coding agents' bash commands]() [Source: berget-ai/guardrails-md]() A small decision model scores every bash command against your GUARDRAILS.md in about 100 ms (maker's figure), before an opencode, pi or Claude Code agent runs it. Commands scoring above the threshold for data loss, secret leaks or rule breaks are blocked; the README warns it can still let risky ones through. - Creator: Berget AI - Stars: 11 - Licence: MIT - Model: berget/bev (System One-compatible) - Tags: security, agent ### Jev as a second pair of eyes for Claude Code [Jev as a second pair of eyes for Claude Code]() [Source: Civalski/claujev]() A Claude Code plugin that asks [Jev]() short typed questions while Claude works: how big is this request, is this command destructive, was the work really done. It suggests an effort level per request, and in enforce mode a confident check can block a "done!" when the work looks incomplete. - Creator: Civalski - Stars: 4 - Licence: MIT - Model: [Jev]() - Tags: coding, agent ### Decision functions inside DuckDB [Decision functions inside DuckDB]() [Source: mraad/duckdb-jev-spatial]() A native DuckDB extension that adds seven Jev-style SQL functions backed by a local [Laya]() model: DuckDB handles geometry and distance, [Laya]() reads the text fields. It uses no hosted model or API keys. - Creator: mraad - Stars: 3 - Licence: Apache-2.0 - Model: [Laya]() - Tags: sql, spatial ### Nimble and Laya play Slay the Spire [Nimble and Laya play Slay the Spire]() [Source: MindrLabs/typed-decision-slay-the-spire]() Runs Bespoke Nimble-9B and three [Laya]() checkpoints as players over the same 200 seeds of a headless Slay the Spire simulator, zero-shot and fine-tuned. The results are the authors' own. - Creator: MindrLabs - Stars: 2 - Licence: MIT - Model: [Nimble](), [Laya]() - Tags: game, benchmark ### Jev DevTools [Jev DevTools]() [Source: diegoag64/jev-devtools]() A drop-in React inspector that turns an app's [Jev]() requests into interactive diagrams of the context, each question and the returned answers. Marked experimental. - Creator: diegoag64 - Stars: 2 - Licence: MIT - Model: [Jev]() - Tags: developer tools, react ### GitHub issues filtered in plain words [GitHub issues filtered in plain words]() [Source: BerriAI/better-backlog]() A local web page and CLI that reads a repo's GitHub issues and lets you filter them by typing a criterion, such as people blocked from upgrading. Each criterion is asked of a decision model through your TypeSafe key or a gateway; add another to narrow the list. - Creator: BerriAI - Stars: 2 - Licence: MIT - Model: [Jev]() - Tags: tooling, web ### Unorphan: loose files filed into their projects [Unorphan: loose files filed into their projects]() [Source: sushrutkhandelwal/unorphan]() A local app you point at a messy folder like Downloads: it works out which of your projects each loose file belongs to, reading the README or CLAUDE.md in your code folders, and files it into a matching folder under your chosen output location. Nothing moves until you press Move, and one click puts it all back. - Creator: sushrutkhandelwal - Stars: 2 - Licence: no licence - Model: [Jev]() - Tags: desktop, files ### Jev and Kev compared on the same questions [Jev and Kev compared on the same questions]() [Source: opper-ai/jev-vs-kev]() Benchmark code and a 362-question test set that sends each item to [Jev]() and to [Kev]() 4B, and reports accuracy, calibration, latency and cost. - Creator: Opper - Stars: 1 - Licence: Apache-2.0 - Model: [Kev](), [Jev]() - Tags: benchmark, evaluation ### Span-01 as a language filter [Span-01 as a language filter]() [Source: sunnydachs/span01-eval]() Tests [Span-01]() Lite as a second check on whether narration text is in the right language, after a regex pass, against a general chat model. The author notes the 23-case test set is small. - Creator: sunnydachs - Stars: 1 - Licence: MIT - Model: [Span-01]() - Tags: evaluation, guardrail ### On-device email triage [On-device email triage]() [Source: patrikcze/mail-triage]() Classifies Apple Mail messages as for you, to review or junk with a [Laya]() model running on your own computer, and shows a report of what it would do. Mail never goes to an external AI service, and it never deletes anything. - Creator: patrikcze - Stars: 1 - Licence: Apache-2.0 - Model: [Laya]() - Tags: email, local ### LinkedIn feed filter [LinkedIn feed filter]() [Source: GuitarWag/linkedin-focus-filter]() A Chrome extension that hides LinkedIn posts that don't match your stated interests, using [Kev]() 0.6B running in the browser on WebGPU. Post text never leaves the browser. - Creator: GuitarWag - Stars: 1 - Licence: MIT - Model: [Kev]() - Tags: browser, extension ### PathJev: decisions on pathology slides [PathJev: decisions on pathology slides]() [Source: CosmoIknosLab/PathJev]() A decision model for histopathology: given a tissue image, a question and candidate answers chosen at run time, including 'none of these', it returns a probability for each in one forward pass. Built on MedGemma-1.5-4B with LoRA adapters and a decision head, with code to rebuild every number in the authors' preprint. - Creator: Cristian Tommasino and Antonio M. Rinaldi - Stars: 1 - Licence: Apache-2.0 (code); weights for non-commercial research only (MedGemma HAI-DEF terms) - Model: PathJev-P (own, MedGemma-1.5-4B base) - Tags: medicine, vision ### MedJev: zero-shot medical decisions [MedJev: zero-shot medical decisions]() [Source: Simonlee711/MedJev]() Runs an open System One model locally on Apple silicon against ten medical benchmarks, from USMLE-style questions to emergency-department admission and 30-day readmission, and compares it with zero-shot LLMs and supervised baselines. - Creator: Simonlee711 - Stars: 1 - Licence: MIT (code); model weights CC BY-NC 4.0 - Model: [OpenJev]() Flash 9B - Tags: medicine, benchmark, local ### Reasoning effort set per message [Reasoning effort set per message]() [Source: prgrmrwy/dsh-auto-model-routing]() A DeepSeek Harness plugin where [Jev](), via OpenRouter, judges each message as simple, standard, complex or escalated and raises or lowers only the reasoning effort of the model you picked. If [Jev]() fails, it switches itself off. - Creator: prgrmrwy - Stars: 1 - Licence: MIT - Model: [Jev]() - Tags: routing, coding ### Ants that draw your picture [Ants that draw your picture]() [Source: scalefreegan/jev-ants]() Each ant is a [Jev]() agent that sees the cells around it and its neighbours' headings and decides which way to move and whether to settle. A thousand small decisions turn a random scatter into a word or an uploaded picture; the maker puts a word at about fifteen seconds and seven cents. - Creator: scalefreegan - Stars: 1 - Licence: MIT - Model: [Jev]() - Tags: art, swarm ### X spam blocker [X spam blocker]() [Source: bit-cook/Jev-X-Blocker]() A Chrome extension that hides spam and noise in X posts and replies by combining [Jev]()'s probabilities for rules you write in plain language with local signals, and decides locally what to hide. Everything hidden can be restored in place, and if the API fails, posts are shown. - Creator: bit-cook - Stars: 1 - Licence: MIT - Model: [Jev]() - Tags: browser, extension ### UI tests written in plain English [UI tests written in plain English]() [Source: andletenkov/decisium]() Write test steps as plain-English Gherkin and [Jev]() picks the element, action and check for each one from options Decisium reads off the page; code, not the model, decides whether a step passed. Resolved steps are recorded so CI can replay them with no model calls. - Creator: andletenkov - Stars: 1 - Licence: MIT - Model: [Jev]() - Tags: testing, playwright ### Noullama, a desktop runtime for local decision models [Noullama, a desktop runtime for local decision models]() [Source: parasite6/noullama]() A desktop app that finds and downloads decision-model GGUFs from Hugging Face, runs them on llama.cpp, and serves a local API with choice, yes/no and score questions, plus a playground for trying presets. - Creator: parasite6 - Stars: 1 - Licence: GPL-3.0 - Model: Local GGUF models - Tags: local, tooling ### RealStop: catch work handed back by a coding agent [RealStop: catch work handed back by a coding agent]() [Source: decision-tune/realstop]() A Claude Code Stop hook with a local fine-tune of [DecisionTune 1.0]() that can block a turn ending that hands work back to the user. VENDOR-REPORTED on 864 test turn ends: the moderate setting caught 67.5% (195 of 289) of hand-backs at 15.8% false interventions, against 56.7% and 14.6% for phrase rules. That test split guided development. Both members were correct in only 16% (37 of 229) of synthetic pairs; the model did not beat phrase rules there. Warm model latency was a median 27 ms on an M5 Pro. The hook cannot verify claimed task completion. - Creator: decision-tune - Stars: 1 - Licence: Apache-2.0 - Model: [DecisionTune 1.0]() (RealStop fine-tune) - Tags: coding, guardrails, local - [Source: Model and evaluation]() ### Offline browser agent on Nimble [Offline browser agent on Nimble]() [Source: jinyoung/nimble-pilot]() A macOS app whose browser agent makes every choice, such as which element to click, as one classification by Nimble-9B running locally in Ollama. Low-confidence calls go to the person. - Creator: jinyoung - Stars: 0 - Licence: Apache-2.0 - Model: [Nimble]() - Tags: browser, agent ### Local intent router with a model cascade [Local intent router with a model cascade]() [Source: indranildchandra/hybrid-intent-router]() Routes support messages through regex, then a small classifier, then decision models, then an LLM as a last resort, all running locally. [Laya]() answers the guardrail and department questions, and [Tev1]() is the default for the next tier. - Creator: Indranil Chandra - Stars: 0 - Licence: MIT - Model: [Laya](), [Tev1]() - Tags: routing, local ### Pacman played by Tev1 in Ollama [Pacman played by Tev1 in Ollama]() [Source: bidouilles/pacman-systemone]() A one-file browser Pacman where the open directions at each step become the options in a single question to a local [Tev1]() model through Ollama's decision endpoint. - Creator: bidouilles - Stars: 0 - Licence: no licence - Model: [Tev1]() - Tags: game, ollama ### Intent router on OpenAI's Decisions API [Intent router on OpenAI's Decisions API]() [Source: anujgupta9860/intent-router-openai-decisions]() A proof of concept that classifies incoming requests with the [Decisions API]() preview and hands them to worker agents. Its README warns that the preview's request format is not yet published. - Creator: anujgupta9860 - Stars: 0 - Licence: no licence - Model: [Decisions API]() - Tags: routing, agent ### Sentence to typed query plan [Sentence to typed query plan]() [Source: saembit/jevquery]() A TypeScript library where [Jev]() picks the table, columns, filter, order and limit from what the schema offers, so it can choose a column but never write a query. Drizzle, Prisma or a SQL builder turns the plan into a real, parameterised query. - Creator: saembit - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: sql, typescript ### Robot boxing: Jev against GPT and Claude [Robot boxing: Jev against GPT and Claude]() [Source: Marcyvan1777/jev-arena]() A 3D boxing game where [Jev](), GPT-4o-mini and Claude Haiku pick each move from the same state and legal actions, and a seeded rules engine resolves the hits. The maker calls the 56-bout results a pilot, not a ranking. - Creator: Marcyvan1777 - Stars: 0 - Licence: no licence - Model: [Jev]() - Tags: game, comparison ### Pet shop catalogue with evidence [Pet shop catalogue with evidence]() [Source: jeromtom/petsona-catalogue-ai]() Tags each pet product's species, life stage and allergens from the store's own text, asking [Jev]() or [Laya]() one typed question per value: supported, contradicted, unknown or conflicting, with the source sentence attached. Merchants can approve, correct and lock tags, or let a validated policy accept them automatically. A hackathon entry for AIBoomi Startup Weekend Kochi, 23-24 Oct. - Creator: Jerom Palimattom Tom (Rexin Dynamics) - Stars: 0 - Licence: MIT - Model: [Jev](), [Laya]() - Tags: e-commerce, shopify ### Daily arXiv digest screened by Jev [Daily arXiv digest screened by Jev]() [Source: Monkeypox966/arxiv-jev-digest]() A GitHub Actions job where [Jev]() scores each new paper from your chosen arXiv categories against each of your research interests, code ranks them and keeps the top few, and only those go to DeepSeek for a summary emailed to you. The default follows humanoid robotics papers. - Creator: Monkeypox966 - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: research, email ### Politicians by Jev [Politicians by Jev]() [Source: n8peace/politiciansbyjev]() Ask a question and [Jev](), via Vercel AI Gateway, gives each of 100 well-known American politicians a yes/no probability; the faces above the threshold float up out of a physics pile. - Creator: n8peace - Stars: 0 - Licence: MIT (code); portraits under their own licences - Model: [Jev]() - Tags: fun, web ### Live interview checklist [Live interview checklist]() [Source: ByronDarlison/call-coach]() A macOS app that transcribes both sides of a call on your Mac with Whisper while [Jev]() ticks off each of your questions once it has been answered, notes which question you just asked, and flags surprising answers for a follow-up. The maker puts [Jev]()'s cost at about 1 to 2 US cents for a 45-minute interview. - Creator: Darlison Holdings - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: meetings, mcp, local ### Voice tutor grounded in a textbook [Voice tutor grounded in a textbook]() [Source: balajibankapur/textbook-voice-tutor]() Answers a seven-year-old's spoken questions from one scanned Grade 2 textbook chapter. [Jev]() decides whether each new question can be answered from the chapter and, with the optional verifier on, whether the drafted answer's claims are supported by the retrieved passages; common questions are matched to cached answers. On the maker's 14-question test, the [decisions API]() cut average answer time from 4.1 to 1.7 seconds and the cost of each gate call about 25 times, against the chat wrapper. - Creator: balajibankapur - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: education, voice ### Job description to occupation code [Job description to occupation code]() [Source: blurboy1985/ssoc-decision-trail]() Paste a job ad and watch [Jev]() walk Singapore's 2024 occupation classification level by level, keeping the three best paths at each step, to a five-digit code, drawn as an animated pixel-art crowd. Six requests, about 1.4 seconds, per the README. - Creator: blurboy1985 - Stars: 0 - Licence: no licence - Model: [Jev]() - Tags: classification, visualisation ### Talk to your Mac in one sentence [Talk to your Mac in one sentence]() [Source: manikv723-jpg/JevVoice]() Speak several instructions in one run-on sentence and [Jev]() picks each action from a menu of skills, and any app from those installed on your Mac, while you are still talking. Anything outside the menu goes to Claude for a proposed shell command, which waits for your approval unless you turn on trust mode. - Creator: manikv723-jpg - Stars: 0 - Licence: no licence - Model: [Jev]() - Tags: voice, macos ### Thunderbird email tagging that learns [Thunderbird email tagging that learns]() [Source: sarath/krestel]() A Thunderbird extension that tags your email with decision models served by Ollama's /v1/systemone endpoint, such as [Tev1]() 0.8B, [Tev1]() 4B or [Nimble](), and keeps learning from the tags you add by hand. - Creator: sarath - Stars: 0 - Licence: MIT - Model: [Tev1](), [Nimble]() - Tags: email, local ### One MCP server for any System One model [One MCP server for any System One model]() [Source: dmonteroh/sys1\_mcp]() An MCP server that gives an AI assistant yes/no, choice and score tools backed by whichever decision model you configure: any /v1/systemone endpoint such as [Jev]() or [Clef](), or [Laya]() running on Apple silicon. Its tools never name a model family. - Creator: dmonteroh - Stars: 0 - Licence: MIT - Model: [Laya](), Jev-Style-0.8B (any /v1/systemone endpoint) - Tags: mcp, local ### A village run on Jev [A village run on Jev]() [Source: Dasge97/the-village]() A pixel-art village in the browser where [Jev](), via OpenRouter, chooses what each adult does for the next two hours, while fixed rules handle urgent needs, night-time sleep and children. The maker built it to test whether [Jev]()'s choices stay sensible and reflect each villager's personality. - Creator: Dasge97 - Stars: 0 - Licence: MIT (code); CC0 1.0 (graphics and sounds) - Model: [Jev]() - Tags: simulation, game ### Build or not: a verdict on your side project [Build or not: a verdict on your side project]() [Source: blakestone-x/buildornot]() Describe a side project and get BUILD, MAYBE or DON'T BUILD with the reasons: [Jev]() rates pain, reach and feasibility out of five and judges whether the GitHub and Hacker News results it finds already do the job. A run costs well under a cent. - Creator: blakestone-x - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: cli, ideas ### PII discovery in plain SQL [PII discovery in plain SQL]() [Source: senthilsweb/jev-pii]() A DuckDB function that asks [Jev]() typed questions about each database column: is it personal data, which kind, and whether GDPR, CCPA-CPRA, HIPAA, LGPD, PIPEDA or DPDP apply. Whether to keep, mask, hash or drop it is decided by code, not the model, and a failed call sends the column to review. Also runs in the browser. - Creator: senthilsweb - Stars: 0 - Licence: Apache-2.0 - Model: [Jev]() - Tags: data, duckdb, compliance ### Search evaluation with Jev as judge [Search evaluation with Jev as judge]() [Source: rickcrawford/jev-relevancy]() Grades search results with [Jev]() on a four-level scale, checked first against Amazon's human-graded shopping queries, and stores runs, results and judgments in DuckDB with the metrics computed in SQL. The author recommends checking it against at least 300 of your own labelled pairs before trusting it on another domain. - Creator: rickcrawford - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: search, evaluation ### decisions, an R package and web app [decisions, an R package and web app]() [Source: fmegahed/decisions-space]() A Hugging Face app built on the R package decisions: upload a set of images or texts, ask [Jev]() one question about each, and get every answer with its probabilities plus the R code to repeat the run. Visitors need no key or account; runs are checked against estimated session and daily budgets. - Creator: fmegahed - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: r, images ### Android notification triage [Android notification triage]() [Source: KoheiTakada/QuietFlow]() An Android app that can use [Jev]() to sort notifications into five classes, from action required to hide, alongside local safety rules and corrections it learns from you. It starts in observation mode and never cancels notifications there; [Jev]() is opt-in with your own key. - Creator: KoheiTakada - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: android, notifications ### West Philippine Sea Watch [West Philippine Sea Watch]() [Source: gianviterbo/wps-watch]() A public monitor of the West Philippine Sea with live ship positions, Sentinel-1 radar passes and Global Fishing Watch events. [Jev]() answers four questions about each item from 13 news feeds, and two of them, whether it serves a Philippine reader and whether it carries PRC state framing, decide whether it is shown or held back. Live at wps.gdgt.link. - Creator: gianviterbo (Gadget Pilipinas) - Stars: 0 - Licence: Proprietary (contact Gadget Pilipinas for reuse) - Model: [Jev]() - Tags: news, maps, monitoring ### Tetris Arena [Tetris Arena]() [Source: arndth/tetrisarena]() [Jev](), SAP-RPT-1, LLMs and a 15-line heuristic play Tetris on the same pieces and candidate moves, with live cost tracking. In the maker's 50-piece race [Jev]() and the heuristic both cleared 19 lines and the others 18, a tie within noise; the difference was speed and cost. - Creator: arndth - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: game, comparison ### Decision models as tool routers, measured [Decision models as tool routers, measured]() [Source: Daewooki/decision-agent-lab]() Calls [Jev](), OpenAI's [Decisions API](), Upstage [Solar Decide]() and local Ollama models through one interface and logs version, latency, tokens and cost. On picking the right one of 12 tools for 40 Korean and English requests, the maker measured [Jev]() 1.13 at 40 of 40 and OpenAI Decisions at 0.90, against 0.85 for a generative model choosing directly. - Creator: Daewooki - Stars: 0 - Licence: MIT - Model: [Jev](), [Decisions API](), [Solar Decide](), [Nimble]() - Tags: agent, benchmark ### Terms of service checker [Terms of service checker]() [Source: shohei81/jev-tos-checker]() A Chrome extension that splits the terms of service on the page into clauses and has [Jev]() check each for eight kinds of one-sided term, from selling your data to forced arbitration, marking only clauses that go well beyond what major services do and showing each next to the original text. The maker says it is not legal advice. - Creator: shohei81 - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: browser, extension, legal ### CI failure triage [CI failure triage]() [Source: tejred213/ci-triage]() [Jev]() reads each failed test and decides the cause: environment or timeout, flaky assertion, genuine regression or API contract mismatch. Code re-runs the build only when every failure would be fixed by re-running (three attempts in all by default), gives flaky tests a low-priority ticket as well, and files tickets for real failures with the test's owner. - Creator: tejred213 - Stars: 0 - Licence: no licence - Model: [Jev]() - Tags: ci, devops ### decide, a CLI and MCP server for Jev [decide, a CLI and MCP server for Jev]() [Source: koltyakov/decide]() A Go command-line tool and MCP server that gives coding agents and shell scripts [Jev]()'s yes/no, choice, score and multi-question calls, with offline request previews and batch runs. - Creator: koltyakov - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: cli, mcp ### Site search that goes to one page [Site search that goes to one page]() [Source: Mou1z/wayfinder]() Visitors type what they need in their own words and [Jev](), via OpenRouter, picks the one page on the site that fits, choosing only from your list of pages. With no key or a slow reply it falls back to word matching. - Creator: Mou1z - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: search, web ### Restaurant order matching [Restaurant order matching]() [Source: prefill-vanishing-point/order-matcher]() Matches messy restaurant orders such as "2 cs portobellos, 5# limes" to exact catalogue items with [Jev]() and flags uncertain lines for a person to check. The maker reports 96% exact matches on held-out test orders. - Creator: prefill-vanishing-point - Stars: 0 - Licence: no licence - Model: [Jev]() - Tags: food, ordering ### Mail server filter that files into folders [Mail server filter that files into folders]() [Source: lesderid/maddy-jev-filter]() Sorts incoming mail for the maddy mail server into IMAP folders: [Jev]() answers questions such as whether a message is spam or phishing, and rules in a config file turn the answers into a folder. It also learns from where you have filed mail before. - Creator: lesderid - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: email, self-hosted ### Which fixes clear most Intune noncompliance [Which fixes clear most Intune noncompliance]() [Source: dallascrilley/intune-why]() Reads Intune compliance data from Microsoft Graph, groups failing devices by the setting that fails, and ranks fixes by how many devices each makes fully compliant. For custom settings its rules don't cover, [Jev]() can optionally pick the remediation path. You can try a live demo with a sample tenant. - Creator: dallascrilley - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: it admin, microsoft ### Classified ads filter [Classified ads filter]() [Source: Aladincz/Jev-Content-Filter]() A browser extension that asks [Jev]() about every ad on the Bazoš classifieds site against criteria you write in plain language, and collapses the ones that don't match. Other marketplaces and auction sites are planned. - Creator: Aladincz - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: browser, extension, shopping ### Ask Jev to flip a coin [Ask Jev to flip a coin]() [Source: royvandewater/jevflip]() A small command-line tool that asks [Jev]() to flip a coin N times and reports the share of heads. - Creator: royvandewater - Stars: 0 - Licence: no licence - Model: [Jev]() - Tags: test, cli ### Hacker News curator on Jev [Hacker News curator on Jev]() [Source: uint4096/hndelegate]() Polls new Hacker News stories every two minutes, runs each unprocessed one through decision trees of [Jev]() questions you define, and sends the stories that match any tree to a Telegram chat or channel. - Creator: uint4096 - Stars: 0 - Licence: GPL-3.0 - Model: [Jev]() - Tags: feed, telegram ### Agent questions answered by Jev [Agent questions answered by Jev]() [Source: jflowers/opencode-auto-question]() An opencode plugin for autonomous runs: when the agent stops to ask the human a question, [Jev]() (or an LLM, if you set one; the default tries [Jev]() first) picks from the offered options. Below a configurable confidence (0.6 by default) it gathers context in a side session and decides again; below 0.35 the human answers after all. - Creator: jflowers - Stars: 0 - Licence: Apache-2.0 - Model: [Jev]() - Tags: agent, coding ### Jev picks which tools an agent loads [Jev picks which tools an agent loads]() [Source: dantetekanem/pi-jev-lazy-load]() A Pi extension that keeps most tool and skill descriptions out of the first request and loads them on demand. In its optional [Jev]() mode, [Jev]() decides before each prompt which ones the request needs. - Creator: dantetekanem - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: agent, cost ### What changes Jev's election forecasts [What changes Jev's election forecasts]() [Source: prashaantr/what-changed-its-mind]() An open research essay on three competitive 2026 US Senate races: compare stored [Jev]() forecasts across different sets of evidence and different orderings of candidates and sources, with the exact requests and saved responses to inspect. The author calls it an experiment, not a calibrated forecast. - Creator: prashaantr - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: research, forecasting ### A decision queue in front of agent tools [A decision queue in front of agent tools]() [Source: iamvazghen/jevqueue]() A Claude Code plugin that checks an agent's tool calls before they run, using hosted [Jev]() when you configure a key and offline rules otherwise. In enforce mode it refuses a call with a reason; advisory mode only reports the verdict. - Creator: iamvazghen - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: agent, guardrails ### Jev for marketing ops [Jev for marketing ops]() [Source: brent-phillips/jev-for-marketers]() A free guide, a Claude skill and four editable recipes for typed marketing-ops decisions: route this lead, tag this asset, is this reply interested. [Jev]() makes the call; an LLM still writes the copy. - Creator: brent-phillips - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: guide, marketing ### Fold low-value posts on X [Fold low-value posts on X]() [Source: Nixz0824/x-semantic-filter]() A Chrome extension: describe what you don't want to see in plain words, and eligible posts on your X timeline get a probability per rule. By default, anything over your threshold folds into a one-line bar you can open again. - Creator: Nixz0824 - Stars: 0 - Licence: MIT - Model: [Jev](), [Clef-omni]() - Tags: extension, feed ### Support tickets routed by Laya [Support tickets routed by Laya]() [Source: ikeremerdem/laya-pet]() A small ticket router: with the [Laya]() backend selected, each support ticket goes to [Laya](), the open Jev-style model, with three questions (which team, how urgent, is it spam). The default backend is a keyword stub, with a slot left for [Jev](). - Creator: ikeremerdem - Stars: 0 - Licence: no licence - Model: [Laya]() - Tags: support, local ### Stock signals with sources, from Parallel [Stock signals with sources, from Parallel]() [Source: hckmstrrahul/parallel-ticker]() A small demo with a live site: type a US or Indian stock and two Parallel Decisions calls, through Parallel's free MCP endpoint, assess its technicals and trading activity from the live web and return probabilities and sources. - Creator: hckmstrrahul - Stars: 0 - Licence: no licence - Model: [Parallel Decisions API]() - Tags: web, trading ### Decision models forecasting Polymarket [Decision models forecasting Polymarket]() [Source: aaron-zhlg/odds-arena]() Keeps the 50 busiest Polymarket questions forecast by several models, including [Jev](), Perplexity's [Decider]() and an OpenAI model, all given filtered web search results but never the price. When markets resolve, each is scored on a leaderboard by Brier score, log loss and accuracy. - Creator: aaron-zhlg - Stars: 0 - Licence: no licence - Model: [Jev](), [Decider]() - Tags: forecasting, benchmark ### Jev Workspace: a mission, some links, a verdict on each [Jev Workspace: a mission, some links, a verdict on each]() [Source: freakngenius/jev-workspace]() Give [Jev]() a mission and some links: it gathers every item, judges each one live with counters and a running API bill on screen, and ends with a results page you can sort, spot-check and export. Runs on your own machine with your own keys. - Creator: freakngenius - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: local, research ### Ask questions about any document, locally [Ask questions about any document, locally]() [Source: nikspence/idec]() A script that asks true/false and multiple-choice questions about an image, PDF, Word file, slide deck, spreadsheet or text, using a decision model on local Ollama. Validated with [d1]():3b and Clef-Flash; [d1]():3b is the default. - Creator: nikspence - Stars: 0 - Licence: MIT - Model: [d1]() - Tags: local, documents ### A 3D disk map that asks Jev what's safe to clear [A 3D disk map that asks Jev what's safe to clear]() [Source: NoxLoveYa/space-sniffer]() A Windows disk-space map that draws your drive as a 3D city and asks [Jev]() which folders look safe to delete; those get a neon outline and land in a Reclaimable list, biggest first. - Creator: NoxLoveYa - Stars: 0 - Licence: GPL-3.0-or-later - Model: [Jev]() - Tags: desktop, windows ### A library's guide: Jev and Claude sort your email [A library's guide: Jev and Claude sort your email]() [Source: byui-library/jev-email-triage]() A step-by-step guide for people who don't write code: set up [Jev]() with Claude Code and build a read-only assistant that ranks the last two weeks of Gmail into piles like Reply needed, Waiting on them and Possible phishing. - Creator: byui-library - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: guide, email ### A moderation queue that learns a new harm in minutes [A moderation queue that learns a new harm in minutes]() [Source: Enayat-Hassani/exposure-aware-routing]() Simulates a moderation queue hit by a wave of privacy-invading posts the existing router did not prioritise: Claude reads 20 of those posts and 20 ordinary ones and drafts one yes/no question, and a fine-tuned version of the open decision model [Laya]() asks it of every new post so the matches jump to the urgent queue. - Creator: Enayat-Hassani - Stars: 0 - Licence: MIT - Model: [Laya]() - Tags: moderation, simulation ### Drex inside DuckDB and other SQL systems [Drex inside DuckDB and other SQL systems]() [Source: nace-ai/drex-duckdb]() Nace's DuckDB extension asks typed questions about rows from SQL and returns probabilities as column values, for routing support tickets, classifying documents or scoring data you already query. Nace also publishes MIT-licensed connectors for Snowflake, PostgreSQL, ClickHouse and Apache Spark, including Databricks SQL. Text and JSON calls default to drex-v2.0; images need an image-capable model such as drex-max. - Creator: Nace AI - Stars: 0 - Licence: MIT - Model: [Drex 2.0](), [Drex Max]() - Tags: sql, database - [Source: Snowflake]() - [Source: PostgreSQL]() - [Source: ClickHouse]() - [Source: Apache Spark and Databricks SQL]() ### Cameo: put yourself in any video [Cameo: put yourself in any video]() [Source: kristopolous/comfyui-cameo]() A ComfyUI workflow that adds you to an existing video. Liquid's d1-2B judges every shot (framing, whether there's room for one more person, where everyone stands) and casts the best ones, then image and video models put you in beside the people already there. - Creator: kristopolous - Stars: 0 - Licence: MIT - Model: [d1]() - Tags: video, comfyui ### Kariyer decision bench: Turkish job ads [Kariyer decision bench: Turkish job ads]() [Source: kubrakrcn/kariyer-decision-bench]() Written at Kariyer.net R&D to pull seniority, department and work type out of job ads. On 600 real ads with 7 questions each (4,200 decisions), [Jev]() 1.13.0 scored 71.5% against 64.6% for the open tau-4b and 65.7% for a prompted 4B LLM. tau-4b was better calibrated (ECE 0.093 against 0.156), and it scored 95.7% on its own training data, which is why the author warns against benchmarking a model on the set it shipped with. - Creator: kubrakrcn - Stars: 0 - Licence: Apache-2.0 - Model: [Jev](), tau-4b - Tags: benchmark, turkish ### Narrative audit: earnings calls against the numbers [Narrative audit: earnings calls against the numbers]() [Source: adityavsb/narrative-audit-jev]() A harness that checks claims from earnings-call transcripts and footnotes against tables in DuckDB, on 64 synthetic cases. Perplexity's pplx-decider-flash-v1.2 scored 85.94% at about 1.4 s a case; a rule baseline scored 68.75% but caught only 2 of 9 Q&A deflections, which the [decider]() caught 9 of 9. A chat model led on accuracy at 90.62%, with confidences stacked at 0.99 to 1.00. OpenRouter's chat endpoint rejected the [decider]() with HTTP 400, so it runs through the separate /api/alpha/decisions endpoint. - Creator: adityavsb - Stars: 0 - Licence: MIT - Model: [Decider]() - Tags: benchmark, finance ### layagrep: train Laya to search code [layagrep: train Laya to search code]() [Source: gak4u/layagrep]() jevgrep finds code by asking [Jev]() thousands of yes/no questions. layagrep records those questions on SWE-bench tasks and fine-tunes the open [Laya]() to answer them locally, with no API key and no code leaving the machine. A smoke run on 400 held-out questions lifted file-navigation AUC from 0.79 to 0.94 ([Jev]() 0.97), but declaration relevance stayed at 0.58 ([Jev]() 1.00). - Creator: gak4u - Stars: 0 - Licence: Apache-2.0 - Model: [Laya](), [Jev]() - Tags: code-search, fine-tuning ### decision-cli: local LLMs as decision models on a Mac [decision-cli: local LLMs as decision models on a Mac]() [Source: cafe3310/decision-cli]() Reads the probabilities of the answer tokens from one forward pass of an ordinary local LLM, with no fine-tuning, and serves them through a CLI, batch mode and a Jev-like API on Apple Silicon. On the 231 public [JevBench]() questions, gemma-4-26b-a4b scored 84.8% at about 170 ms a question and ling-3.0-tiny 69.7% at about 45 ms; most of the misses are on the hard set. - Creator: cafe3310 - Stars: 0 - Licence: Apache-2.0 - Model: Local MLX models - Tags: cli, mlx, local ### desktop-agent: Windows apps run in the background [desktop-agent: Windows apps run in the background]() [Source: elidot007/desktop-agent]() Lets any agent operate Windows desktop apps without taking over the screen. A local 0.33B decision model, S1-Fast, reads the app's accessibility element table and does the sure steps. Uncertain steps go back to the agent or to a fine-tuned OneJev-4B, and irreversible ones stop until a person confirms them. It reports 68.4% macro step success on five public held-out test sets and 40 of 40 on its own development tasks. - Creator: elidot007 - Stars: 0 - Licence: Apache-2.0 - Model: S1-Fast (own 0.33B), OneJev-4B - Tags: desktop, agent, windows ### Decision model kernels: a faster llama.cpp on one GPU [Decision model kernels: a faster llama.cpp on one GPU]() [Source: WyrmKeep/decision-model-kernels]() A lab notebook for speeding up one decision model, H2O Lightning 4B, served through llama.cpp on an RTX 3060 Ti. Each request takes 2.13 to 2.45 times its roofline floor. The first measured change, a llama.cpp pull request, cut latency by 1.94% with bit-for-bit identical answers; the bigger gains in its list of 15 targets are labelled predictions. - Creator: WyrmKeep - Stars: 0 - Licence: MIT - Model: [H2O-Lightning-4B]() - Tags: inference, llama.cpp, cuda ### JEV Damocles: when a group-chat bot should speak [JEV Damocles: when a group-chat bot should speak]() [Source: ch7flyer/astrbot\_plugin\_jev\_damocles]() An AstrBot plugin that decides whether a bot in a group chat should reply. It merges bursts of messages, separates conversation threads, then asks [Jev]() four questions, batched into one request when the backend supports it (was the bot addressed, does it follow, should it reply, how urgent). For messages without a direct mention, quote-reply or poke, it uses a high and a low reply threshold alongside other checks, and scores in between can be reassessed later; those direct interactions can skip the gate. The README suggests a 3 to 7 day dry run first. - Creator: ch7flyer - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: chat, bot, astrbot ### jev-router: Jev sets Claude Code's reasoning effort [jev-router: Jev sets Claude Code's reasoning effort]() [Source: pyxl-dev/jev-router]() A Claude Code plugin that asks [Jev]() one choice question per message, low, medium or high effort, and applies the answer when [Jev]() is at least 60% sure. It changes effort, never the model, so the prompt cache survives. Subagents are never routed. - Creator: pyxl-dev - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: claude-code, plugin, routing ### JevRouter: Jev picks the next step, a policy decides [JevRouter: Jev picks the next step, a policy decides]() [Source: HangChi/JevRouter]() An agent runtime where an LLM reads the task, one [Jev]() call answers four typed questions (next action, risk, whether to confirm, whether to reason further) and a local policy engine decides what may run. On 53 versioned test cases it hit the expected action 92.5% of the time against 96.2% for an LLM baseline, at about 1.0 s against 4.0 s, and asked for confirmation on 16 of the 17 cases that needed it. - Creator: HangChi - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: agent, runtime, policy ### skill-tester: grading agent skills with Jev [skill-tester: grading agent skills with Jev]() [Source: milesroxas/skill-tester]() An eval harness for agent skills: headless agent runs, code checks, [Jev]() judge packs and blind human review; the generic pack asks one yes/no question per skill rule. Graded with the writing pack, the reference skill's v2 run matched its original benchmark's report within 2 points on every layer and in total. - Creator: milesroxas - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: evals, agent-skills ### dsh-cu-d1: Liquid's d1-2B drives a browser and a Mac [dsh-cu-d1: Liquid's d1-2B drives a browser and a Mac]() [Source: Alyosha28/dsh-cu-d1]() Uses Liquid's open d1-2B as a local System 1 for DeepSeek Harness Browser Use and macOS Computer Use, with approvals and routing. Downscaling wide screenshots before the model reads them cut one decision from 879 ms to 310 ms on a 2508-pixel window, and six consecutive trials all chose a click action. d1-2B itself is under the LFM Open License, not MIT. - Creator: Alyosha28 - Stars: 0 - Licence: MIT - Model: [d1]() - Tags: computer-use, browser, local ### Agastya: walking a document's outline to find answers [Agastya: walking a document's outline to find answers]() [Source: cars3/agastya]() Retrieval that walks a document's table of contents, asking [Jev]() one yes/no question per section, then ranks the sections it kept with EmbeddingGemma. On all 6,005 test queries of the public SearchTome release it returned 89.1% of gold sections at 3.6 sections per query, against 84.5% at 3 for EmbeddingGemma alone; its top-1 and top-3 recall equal the embedding model's. [Jev]() alone, ranking flat, scored 54.2 top-1. It needs no training and runs on one 8GB consumer GPU, and failed upgrades are published too. - Creator: cars3 - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: retrieval, search ### Jev picks the move for a robot arm [Jev picks the move for a robot arm]() [Source: uhiron55/physical-ai-2026-advanced-1-final-assignment]() A course project where [Jev]() chooses the transfer strategy for a CRANE+ arm in a Gazebo simulation, MoveIt 2 plans the path and ROS 2 controllers run it. In three episodes on authenticated [Jev]() decisions with no fallback, [Jev]() picked the home transfer at 0.94 to 0.96 confidence and all three placed the cube, 2.70 to 7.48 mm off target. These are scripted-teacher episodes; the README's separate learned-policy (ACT) tests are not reported as grasp success. - Creator: uhiron55 - Stars: 0 - Licence: Apache-2.0 - Model: [Jev]() - Tags: robotics, ros2 ### jev-droid: Android automation with two models [jev-droid: Android automation with two models]() [Source: frederico-kluser/jev-droid]() Drives an Android emulator from plain language: DeepSeek V4.1 Flash decides the next step, and [Jev]() triages each screen (loading, error, permission prompt) and guards actions before they run. The three-step demo in the README, Settings to Network to Wi-Fi, cost $0.00019 and took 21 seconds; every step saves its screen, decision, tokens and cost. - Creator: frederico-kluser - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: android, agent ### topic-gate: reset the chat when the topic changes [topic-gate: reset the chat when the topic changes]() [Source: tauseefau/topic-gate]() Asks [Jev]() one question per message, 300 to 600 input tokens, to tell whether the conversation is on the same topic, a subtopic or a new one, and recommends a session reset only when a new topic wins clearly (p of at least 0.65 and 0.25 ahead), so you stop paying to resend a dead transcript. Checked against live [Jev]() on related, unrelated and short messages. - Creator: tauseefau - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: context, agent ### Laya Notes: notes that tag themselves [Laya Notes: notes that tag themselves]() [Source: itsdula/laya-notes]() A desktop notes app for Mac, Windows and Linux that keeps plain Markdown files and runs the open [Laya]() model on your own computer. Each tag is a yes/no question, like "Is this note about a trip?"; [Laya]() answers all of them for a note in one pass and keeps the tags it is confident about. Notes stay on the computer; the only network request is a one-time download of about 1.6 GB of model files. - Creator: itsdula - Stars: 0 - Licence: MIT - Model: [Laya]() - Tags: notes, desktop, local ### pi-omarchy-cu: Linux desktop control for the Pi agent [pi-omarchy-cu: Linux desktop control for the Pi agent]() [Source: Paul-Yuchao-Dong/pi-omarchy-cu]() Lets the Pi coding agent drive real Linux desktop apps on Omarchy (Hyprland Wayland) by reading the accessibility tree and asking [Jev]() for the target element and the operation. It offers 11 Wayland-safe element operations. Jev-driven actions abstain below 0.55 confidence and need confirmation below 0.70, or 0.90 for destructive ones; direct operations on a known element run without [Jev](). - Creator: Paul-Yuchao-Dong - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: computer-use, linux, agent ### Decision Wall: 200 questions about one text [Decision Wall: 200 questions about one text]() [Source: charles-forsyth/jev-decision-wall]() Asks [Jev]() 200 yes/no questions about one text in a single call and shows the answers as a heatmap. The keyless hosted demo replays recorded answers. VENDOR-REPORTED: about 185 ms for one question, 200 ms for 50, and 150 to 300 ms for 200; seven parallel calls returned 1,400 answers in 282 ms, against 1,092 ms back to back. The author measured [Jev]() 1.14 from a US home connection and has not measured calibration. - Creator: charles-forsyth - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: classification, heatmap ### Jev-Router: choose a model after applying hard constraints [Jev-Router: choose a model after applying hard constraints]() [Source: AIAnytime/Jev-Router]() An OpenAI-compatible router that filters models by hard constraints before [Jev]() judges which one fits. VENDOR-REPORTED, on its own 30 prompts: tier accuracy was 100% against 53.3% for a heuristic, task-type accuracy 83.3% against 46.7%, and hard prompts sent to a weak model 0 against 14. Estimated savings against always using the strongest model, from token counts and configured prices rather than running the chosen models, were 35.6% against 68.7%; the heuristic's estimate was higher because it under-routed. Decision latency was 295 ms at p50 and 409 ms at p95. - Creator: AIAnytime - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: routing, agent ### Laya and ZCode: local checks and context selection [Laya and ZCode: local checks and context selection]() [Source: Ydjin1984/laya-zcode-stack]() Runs multilingual [Laya]() through a local Go/ONNX service, with a ZCode plugin for risk checks, context selection and 13 MCP tools. VENDOR-REPORTED on an RTX 3060, three runs per variant: total input fell from 359,085 to 239,544 tokens (33.3%) and requests from 16 to 9, but non-cached input rose 21.8%. [Laya]() ranking alone cut total input 9.3% and non-cached input about 2.8%, within the observed range; the author has not established a fresh-token or cost saving. The /decide endpoint took about 26 ms at p50. - Creator: Ydjin1984 - Stars: 0 - Licence: MIT - Model: [Laya]() - Tags: coding, context, local - [Source: Benchmark and caveats]() ### PR wall: separate judgments about each pull request [PR wall: separate judgments about each pull request]() [Source: House-elves/pr-wall]() Puts open pull requests on a wall and asks [Jev]() eight typed questions about each diff and its tests, using quarkusio/quarkus by default. It keeps the answers separate and never combines them into a score. Right/Wrong feedback records corrections and groups accuracy by confidence as you label answers; no accuracy figures are published yet. Its questions and Quarkus context are adapted from brunobat/quarkus-pr-wall. - Creator: House-elves - Stars: 0 - Licence: Apache-2.0 - Model: [Jev]() - Tags: code-review, calibration ### openclaw-d1-local: local decisions for OpenClaw [openclaw-d1-local: local decisions for OpenClaw]() [Source: heff0/Kev-claw-decisionmaking]() Supplies OpenClaw with local typed decisions from LiquidAI/d1-Omni-2B-MLX-8bit on Apple Silicon. It has a native decision provider and a bridge for OpenClaw's TypeSafe integration. The repository name contains [Kev](), but the runner loads d1-Omni. VENDOR-REPORTED on an M4 Mac mini: about 200 ms per warm batch and about 3 GB of resident RAM. - Creator: heff0 - Stars: 0 - Licence: MIT - Model: [d1]() - Tags: agent, local, apple-silicon ### inspect-drex: score agent outputs against a rubric [inspect-drex: score agent outputs against a rubric]() [Source: nace-ai/inspect-drex]() A scorer for the UK AI Security Institute's Inspect framework that asks [Drex]() whether an output meets your criteria and keeps the probabilities, so you can change pass thresholds after the run. It defaults to drex-max, whose target changes with flagship releases; pin a model for repeatable runs. [Drex Max]() accepts images, while drex-v2.0 accepts text and JSON. - Creator: nace-ai - Stars: 0 - Licence: MIT - Model: [Drex 2.0](), [Drex Max]() - Tags: evaluation, agent ### OwnYourTimeline: filter the X Home feed [OwnYourTimeline: filter the X Home feed]() [Source: raghavio/ownyourtimeline]() Filters the For you and Following feeds in Chrome or Edge using a plain-language instruction and [Jev](). It sends one bulk TypeSafe request per timeline response, splitting to stay within input budgets, and removes posts scoring at least 0.90 before X renders them. Errors and timeouts leave posts visible; it checks at most 80 posts per response. The built extension has Chromium tests with X and TypeSafe responses stubbed. No filtering-accuracy figures are published. - Creator: Raghav Sharma - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: feed-filtering, browser-extension ### Jev router for GitHub Copilot [Jev router for GitHub Copilot]() [Source: TechPreacher/jev-router-copilot]() [Jev]() sizes a Copilot message as tiny, everyday, large or hardest and sends it to a helper agent pinned to a model. Defaults are Haiku 5.5, Sonnet 5.5 and Opus 5.5, with the hardest helper configured to request high effort. Recorded helper runs used medium effort. The author records live runs of the hook in Copilot CLI 1.0.105 and VS Code 1.144, plus the CLI wrapper. The MCP fallback is tested offline only. Its 54 offline tests passed when checked on 03 Nov 2026; it publishes no savings benchmark. - Creator: Sascha Corti (TechPreacher) - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: model-routing, coding-assistant - [Source: Live CLI wrapper run]() - [Source: Live VS Code run]() ### clef-playa [clef-playa]() [Source: wevans997-dev/clef-playa]() [Clef]() scores tiles from Jamen Percy's 2026 Burning Man gigapan, then the code sums expected RV counts, car counts and shade coverage. The author's 256 px run judged 8,785 tiles for $5.29, estimating 14,385 RVs, 18,084 cars and 256.8 acres of shade. These are estimates: vehicles under shade are hidden, and dense 512 px tiles undercount. Splitting validation tiles into 256 px quadrants brought RV estimates to 386 against 394 from the LLM counters. - Creator: wevans997-dev - Stars: 0 - Licence: MIT - Model: [Clef]() - Tags: aerial-imagery, counting Author's recorded runs · VENDOR-REPORTED ### dsh-system-one [dsh-system-one]() [Source: Xunzhuo/dsh-system-one-plugin]() A DeepSeek Harness plugin uses [vLLM-SR d3-ultra]() to triage steps, guard tool calls, strip injected lines and check whether work is done. This is the model maker's integration, by Xunzhuo Liu, co-founder of vLLM Semantic Router. In his small probes, the guard got 35 of 36 commands right, stopping all 16 risky ones, and the shield got 22 of 22 passages right. He reports a 0.49 s median across 53 guard decisions. Triage of 230 tickets took 121 s and 117,877 LLM prompt tokens, against 543 s and at least 510,986 for the LLM alone. - Creator: Xunzhuo Liu - Stars: 0 - Licence: Apache-2.0 - Model: [vLLM-SR d3-ultra]() - Tags: agent-guardrails, coding-assistant Author's recorded runs · VENDOR-REPORTED - [Source: Maker profile]() - [Source: Maker's website]() ### mail-agent [mail-agent]() [Source: Jaydccq/mail-agent]() Incoming Gmail goes through [Jev]() triage before alerts reach Telegram or Lark. Claude or Codex CLI can draft replies, which wait for a Send tap. One Python process uses one SQLite file. The author's testing notes record real [Jev]() calls on nine synthetic emails at about 0.1 s each, plus an end-to-end run with real Gmail, Telegram, [Jev]() and Codex CLI; they report no triage accuracy figure. - Creator: Jaydccq - Stars: 0 - Licence: Apache-2.0 - Model: [Jev]() - Tags: email-triage, notifications Author's recorded runs · VENDOR-REPORTED - [Source: Testing notes]() ### Sluice [Sluice]() [Source: Potterluo/sluice]() A coding harness lets a worker propose state transitions and a judge approve, reject or redirect them. Its 02 Nov 2026 real-model run used [Jev]() 1.14 as judge and deepseek-v4-pro as worker. The author reports five judge calls for under $0.0003; the allowed-choice menu corrected three requests for nonexistent states. A separate five-task SWE-bench pilot does not identify its judge, so its score cannot be attributed to [Jev](). - Creator: Potterluo - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: coding-assistant, agent-controller Author's recorded runs · VENDOR-REPORTED ### Drone Commander, Powered by Jev [Drone Commander, Powered by Jev]() [Source: Ramolisdenneyous/Drone-Commander-Powered-by-Jev]() In this local field test, [Jev]() picks a unit's action from legal option IDs supplied by a tactics game's rules engine. The server checks the choice; a deterministic fallback handles late, malformed or weak replies. The author estimates about 75% lower token cost for those decisions with no obvious loss of tactical ability. Those are early playtest impressions. The first matches set both confidence and margin gates to zero, letting every legal [Jev]() choice play. - Creator: Ramolisdenneyous - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: games, tactics Author's recorded runs · VENDOR-REPORTED ### Chinese game-comment analysis [Chinese game-comment analysis]() [Source: WanKun824/pbz-comment-insight]() A local laya-multilingual fine-tune labels 9,696 Bilibili comments about Phantom Blade Zero with 14 questions and a two-level topic tree. The author reports 60.8 minutes of training on an RTX 3060. On 771 test comments with 1,485 decisions, topic accuracy rises from 0.210 to 0.911 and validity from 0.253 to 0.971. Test labels come from rules, whose topic precision is estimated at 72% to 75%, so 0.911 measures agreement with the rules. The README also discloses some train/test overlap. Toxicity detection fails at macro-F1 0.489, with only 11 training positives. - Creator: WanKun824 - Stars: 0 - Licence: Apache-2.0 - Model: [Laya]() - Tags: comments, classification Authors' published runs · VENDOR-REPORTED; not independently reproduced - [Source: Apache-2.0 licence]() ### GitHub issue labels and priorities [GitHub issue labels and priorities]() [Source: elcann7/jev-issue-triage]() A GitHub Action asks [Jev]() nine questions per issue in one request, then a decision table turns the probabilities into labels and priority tiers. Probabilities between 0.3 and 0.7 count as unsure. It has a dry-run mode. The author's run passes 65 of 66 checks on 16 synthetic issues for about $0.001. That small synthetic set includes a fixture revised after its fake-example wording affected the answer. - Creator: Elcan Səfərli - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: issues, automation Authors' published runs · VENDOR-REPORTED; not independently reproduced - [Source: MIT licence]() ### Prompt feedback while typing [Prompt feedback while typing]() [Source: filip-a-9/jev-prompt]() This Claude Code mod asks [Jev]() through OpenRouter to grade the prompt being typed and name what it lacks. Hints are advisory; zero-data-retention routing is on by default. On 84 prompts, the band agrees with labels 71% of the time against an 80% target. Across 217 prompts with known outcomes, hints appeared on 25; 24% of those were followed by trouble (95% interval 11% to 43%), against a 10% overall rate. AUC is 0.60. Three blind Claude labellers supplied the labels. The harness is private, and whether the hints improve prompts has not been shown. - Creator: filip-a-9 - Stars: 0 - Licence: MIT - Model: [Jev]() - Tags: coding, prompts Authors' published runs · VENDOR-REPORTED; not independently reproduced - [Source: MIT licence]() ## What independent research has found so far Papers, leaderboards, head-to-head tests, a security test and a hands-on review, then a small test we ran ourselves. ### Alert paging · authors' numbers · VENDOR-REPORTED Coroot replays 31 synthetic p99-latency scenarios minute by minute, with a three-minute hold before paging. Its burn-rate rule gets 30 of 31 right; [Jev]() 1.13 tuned at 0.59 gets 29; [Tev1-4B]() at 0.5 gets 28; untuned [Jev]() and tuned [Nimble]() 9B get 27. Explaining the hold raises [Tev1-4B]() from 23 to 28. Coroot reports about $30 for two OpenRouter passes, including eight cents for [Jev](). Thresholds were tuned on these same scenarios, designed to be obvious to a human with a fixed 1 s SLO. [Coroot · Anomaly detection for alerting · 03 Nov 2026]() ### Issue calibration · author's numbers · VENDOR-REPORTED On 2,319 GitHub issues with one human type label each, typesafe-cascade reports 76.7% accuracy for [Jev]() 1.14 with no threshold. At 0.8 it keeps 1,964 issues at 81.9% accuracy; at 0.95 it keeps 1,689 at 85.2%. The 1,794 issues in the 0.9 to 1 confidence bucket average 98.9% confidence but only 84.3% accuracy. Excluding question labels gives 99.1% confidence and 91.7% accuracy in that bucket. The author suspects maintainers' resolved-issue labels add noise and would fix the label set before tuning the threshold. The sample is stratified by label, unlike a real backlog. [monumei · typesafe-cascade · Jev 1.14.0 results · 03 Nov 2026]() ### Weight patch · author's numbers · VENDOR-REPORTED A lowercase positive predicate makes the English [Laya]() root checkpoint answer false at full confidence regardless of the state in code2saurabh's reproduction of issue #1244. Rank-1 edits to layers 16 to 21 use 12,288 learned numbers, a 54 KB patch fitted in about six minutes on a GTX 1650 Ti. On 872 SST-2 validation reviews, accuracy rises from 0.491 to 0.888 and AUROC from 0.496 to 0.940; wrong answers at confidence at least 0.9 fall from 443 to 13. Two held-out wordings score 0.858 and 0.890. The labels workaround still scores higher at 0.908. Only the English root checkpoint was tested; the negative wording stays near 0.62. [code2saurabh · Laya #1244 weight patch · English checkpoint 7b928d82, laya 0.4.21 · 03 Nov 2026]() ### LLM judges · arXiv Across eleven judge benchmarks, Subramanian tests 14 decision models against generative judges. Five are statistically indistinguishable from GPT-5.6 without reasoning across all five pairwise and best-of-4 benchmarks. With a reference answer, eleven reach 90.3% to 93.4% on JudgeBench, statistically indistinguishable from Opus 5.5 and GPT-5.6 with reasoning. Injection reverses 0% to 5% of correct verdicts for five models, but 17% to 49% for the other nine. The figures are Subramanian's measurements. [Shreyas Subramanian · To Jev or Not to Jev · arXiv 2611.00505 · 31 Oct 2026]() ### Languages · arXiv An English winner need not win elsewhere. XDecideBench covers nine suites, 122 languages, 88,425 texts and 153,875 typed questions. The English-best model, chosen in hindsight, loses significantly in 56 of 232 suite-language cells. A confidence threshold fitted on English for 5% errors admits 7.8% to 12.4% on German and French XNLI. These are Šuppa and Kopčan's measurements; thresholds need checking in the language and task where they will run. [Marek Šuppa, Jaroslav Kopčan · XDecideBench · arXiv 2611.01370 · 01 Nov 2026]() ### Policy gates · authors' numbers · VENDOR-REPORTED Examples and counterexamples changed how policy gates behaved in agentapprove's study. James P. Beno tested 22 hosted decision models and two prompted general models on 2,643 synthetic tool calls labelled blind against 28 rules, with a sealed holdout. Explicit boundaries raised recall by 5.5 to 26.1 points on all 23 models still available and reduced false positives on 21. [Jev]() 1.14's recall rose from 89.3% to 98.2%; unnecessary interruptions fell from 30.6% to 4.8% of 252 clean actions. With examples, its median latency was 120 ms at $0.46 per thousand requests. The author sells the policy engine. [James P. Beno, agentapprove · Decision Models as Policy Gates for Coding Agents · arXiv 2611.02155 · 01 Nov 2026]() ### Cognitive tasks · arXiv S1CogBench tests 24 cognitive tasks in 27 modes. [Laya]() fails to beat uniform choice on 17 of 26 targets, against 3 for [Intern-Decision-4B](), which beats its Qwen3.5 backbone on 23. In Mouselab, Qwen earns the highest autonomous reward but predicts humans worse than uniform choice; [Laya]() fits humans best among the open models yet earns zero reward. Under the same readout, [Intern-Decision]() is no faster than its same-size backbone. Generating text takes 80 to 316 times as long as one forward pass in the two general models timed. Hou and Yamada report these results. [Dengzhe Hou, Kazunori D. Yamada · S1CogBench · arXiv 2611.00911 · 31 Oct 2026]() ### Vulnerability triage · authors' numbers · VENDOR-REPORTED At a fixed review budget, Cyber-JEV misses fewer vulnerable functions in its authors' test. An ensemble of five fine-tuned Qwen3.5-4B models reaches AUROC 0.814 on 7,270 functions, against 0.743 for [Jev]() 1.13. With both limited to 4,031 alerts, it misses 16 of 322 vulnerable functions instead of 39. The method trains on matched vulnerable, patched and unchanged functions; the paper links no code. [Zhao et al. · Cyber-JEV · arXiv 2611.02957 · 02 Nov 2026]() ### Live-web benchmark · vendor's own · VENDOR-REPORTED Parallel's PWS-Decisions averages three runs of 300 questions on 29 Oct 2026 (Pacific). Its API with live web search scored 76.22%; Perplexity [Decider]() v1.2 scored 50.89%, OpenAI's [Decisions API]() 41.67% and [Jev]() 1.14 33.44% without web access or supplied external context. Parallel puts uniform random guessing at 42.42%. GPT-6 Luna and Claude Haiku 5.5 with native search scored 97.33% and 96.56%, at $16.52 and $27.43 to $28.79 per 1,000 questions, against Parallel's $1.00; median latency was 7.42 and 6.38 seconds against 0.56. Parallel generated synthetic questions from its own customer topics and search traffic, with separately researched answers and exact-match scoring; refusals count as wrong. It calls this a development challenge set: 100 questions are gated, 200 private, and questions, answers and per-question predictions cannot be redistributed. The runner is MIT. The blog chart labels Baseten Elephant at 32.4% overall and 32.3% public; the repo instead records Mammoth at 45.0% and 37.0%. [Parallel · PWS-Decisions v0.1.0-rc1 · blog, repository results and dataset terms · 01 Nov 2026]() ### Retrieval routing · independent · author's numbers Adding Clef-Flash 9B as a domain router reduced pooled nDCG@10 from 0.4503 to 0.4084 in nicklisac's retrieval experiment. Per-domain LoRAs with the router reached 0.4923; a joint LoRA without it reached 0.4947. Perfect domain assignment to specialists gave an oracle ceiling of 0.4991. The detailed report and logs use 2,161 held-out queries over FiQA, ArguAna, SciFact and CQADupstack-gaming, with EmbeddingGemma 2 on one RTX 3090; the README says 1,081. Post-retrieval reranking and pruning with Clef-Flash also reduced the corresponding embedding-only scores. [nicklisac · lora-router-rag · REPORT.md and results/run\_log.jsonl · checked 03 Nov 2026]() ### Untuned 4B baseline · independent; Together comparison · VENDOR-REPORTED Using [Tev1]()'s own decision protocol and v1 evaluation data through local Ollama, jev-rsi's author measured untuned Qwen3.5-4B at 673/1,000 (67.3%) on test and 130/300 (43.3%) on policy\_transfer. Together reports 880/1,000 (88.0%) and 300/300 (100%) on evaluation sets used during development (VENDOR-REPORTED). In a separate frozen-backbone experiment with small heads and a fixed self-improvement loop, test scores ranged from 6.7% to 31.1%, against a 29.8% random baseline, on a small held-out set. The author concludes that heads alone did not learn the decision task in this setup. [Joshhuang123 · jev-rsi · baseline and frozen-head experiments · 02 Nov 2026]() ### Earth observation · authors' numbers · VENDOR-REPORTED EO-Jev adapts d3-flash (8.4B, Qwen3.5-9B backbone), from the [Decision 1.0]() makers, with LoRA for Earth observation. Across 30 settings from 23 public benchmarks, its authors report significant improvements in 26 settings, averaging +23.8 accuracy points in-domain and +8.0 on 12 held-out settings. It matched or exceeded the best open EO specialist in 20 of 28 settings and had lower calibration error in 21 of 28. Median throughput was 6.4x the specialist's with each at its best batch size, or 3.2x with one question per call. General decision accuracy fell 1.8 points. Counting, change quantification and calibration far from the training data remain weak spots. The repository has code and reports; it excludes weights and run logs, and lists the paper as coming soon. [Abouelhadid, Abuhani, Munir, Khan, Khan · MBZUAI and collaborators · EO-Jev · 02 Nov 2026]() ### Defect detection · authors' numbers · VENDOR-REPORTED With two reference images per class across five manufacturing case studies, [Jev]() scored 0.75 to 1.00 test accuracy and Clef-Flash 0.71 to 0.99, against CLIP's 0.67 to 0.84 using the same references. The authors' agentic description refinement improved accuracy in six of eight model/application combinations, and their calibration reduced Brier score in 13 of 15 configurations. [Megahed, Chen · Beyond CLIP · arXiv 2610.41199 · 30 Oct 2026]() ### Visual inspection · authors' numbers · VENDOR-REPORTED VerdictAD uses Clef-Flash 9B to refine existing anomaly detectors with decisions about whole images and local regions. In the authors' four-shot tests, averaged over three reference draws, PatchCore's image AUROC rose from 93.0 to 97.5 on MVTec AD, 89.0 to 93.9 on VisA and 67.4 to 78.0 on MVTec LOCO AD. DINOv2-L rose from 68.7 to 77.5 on LOCO. These are image-level AUROC figures, not accuracy or pixel-level scores; the repository includes the paper and experiment logs. [Sun et al. · VerdictAD · code, paper and logs · 02 Nov 2026]() ### Training case study · authors' numbers · VENDOR-REPORTED One developer trained a decision head and LoRA on Qwen3.5-0.8B with Unsloth to screen job postings. On 63 receipt-labelled test examples, accuracy was 0.762 and macro-F1 0.683, against 0.714 and 0.278 for always rejecting. On a separate 460-posting check against an agent's historical verdicts, accuracy was 0.726, below the 0.835 always-reject baseline. That external check uses a different population and reference. The author's data and trained weights stay private, so a public clone cannot reproduce the scores. [jayvir-git · training-own-decisions-model · 02 Nov 2026]() ### Medical fine-tune · authors' numbers · VENDOR-REPORTED A medical LoRA on Qwen3.5-4B scored 67.90% on 1,514 external questions, against 59.91% for the base with the same scoring method and 70.87% for archived [Jev]() 1.13. The full-set score counts 100 questions with 148 options as wrong because the released interface supports at most 52. On 100 triage questions it answered 40 correctly, against [Jev]()'s 66. The maker has not scored 1,620 sealed questions; training labels lack individual clinical-expert review and the probabilities are uncalibrated. [JuneYao · medical-jev-4b · model card · 02 Nov 2026]() ### Pathology · authors' numbers · VENDOR-REPORTED PathoJev encodes a pathology slide once and answers several questions over that shared visual context. Across three whole-slide benchmarks, the authors report mean accuracy of 77.2% for 9B and 78.1% for 27B, exceeding SlideChat by 5.4 and 6.3 percentage points. The 9B model ran 9.1 times faster than SlideChat on WSI-Bench. Against encoding the prefix separately, shared context was 6.9 times faster for eight questions and 11.2 times for 64 at 9B; the 27B figures were 5.9 and 8.9 times. [Zhang et al. · PathoJev · arXiv 2610.40752 · 29 Oct 2026]() ### Search reranking · authors' numbers · VENDOR-REPORTED DecisionRank trains a decision reranker by having a generative teacher judge ranked lists built from the student's scores. The authors report that a 1.7B student beat its 27B generative teacher on TREC DL19 to DL23. At 8B, it beat its generative listwise counterpart while running 24.5 times faster on TREC DL and 16.7 times faster on BRIGHT. [Wu et al. · DecisionRank · arXiv 2610.42128 · 30 Oct 2026]() ### Agent controller · authors' numbers · VENDOR-REPORTED Operon-S1 turns recurring judgments in a biomedical agent into closed questions at 13 decision points. A local decision model answered in-loop requests at a median 0.6 seconds on CPU, with no measurable agent-score change on five of six benchmarks. Asking the driving LLM those same questions instead accounted for 8.0% to 35.2% of task cost. Genoria AI's study spans seven life-science benchmarks overall; the unchanged-score comparison covers six, and task-level results use the public checkpoint. [Hou et al., Genoria AI · Operon-S1 · arXiv 2610.41350 · 30 Oct 2026]() ### Benchmark · independent Liquid's open 2B decision models against two 2B chat models (Qwen3.5-2B, InternVL3.5-2B) on 248 real pictures, all on one RTX 5060 Ti. d1-2B routed scanned documents to one of 16 types at 88.8% against 55% and 56%; d1-Omni-2B classified 71 of 73 clear frames correctly as violent or not (97.3%); on a 50/50 defective-part set the [d1]() models scored 70.8%. [d1]() was 1.6 to 23 times faster per question. The author notes it is not quite like for like: the chat models reply in text. On 500 text items d1-2B agreed with [Jev]()'s answers 68% of the time. [Aditya (@0xironyAditya) · d1-2B vs three 2B models · 01 Nov 2026]() ### Benchmark · independent [Jev]() as a search reranker against Cohere, Voyage and Qwen rerankers on multilingual MIRACL and Japanese JQaRA. Scoring each passage with its own yes/no question, batched per query, [Jev]() 1.14 scored nDCG@10 0.7496 on MIRACL and 0.8849 on Japanese JQaRA, the best mean of the final seven (0.8173, against 0.8061 for Cohere rerank-4-pro), at $0.89 to $1.14 per 1,000 queries against Cohere's $2.50. In a separate 252-query MIRACL run most dedicated rerankers beat [Jev]() 1.13 by a significant margin; results differ by dataset, model version and scoring method. [Jeffrey (Dongkyu) Kim · jev-reranker · 01 Nov 2026]() ### Game test · independent Decision models playing Super Mario Bros. world 1-1, one choice per step. From the game state, Perplexity's [Decider]() v1.2 cleared the level in 20.13 game seconds for $0.0078, and Liquid's d1-ultra and Inception's [Mercury Decide]() also cleared it. From pixels alone, [Decider]() v1.2 and Baseten's [Mammoth 1]() cleared it; [Jev]() 1.14's best pixel run reached 51.94% of the level. These are 46 development runs with different settings, which the author says are not a ranking. [Nicholas Greenwood · mario-bench · 01 Nov 2026]() ### Benchmark · independent On the 127 test documents of the Text Anonymization Benchmark (European Court of Human Rights cases), Cloudflare's extraction model [Clairic]() flagged more identities than OpenAI's Privacy Filter but was less precise: 77.7% of its spans were ones annotators marked to hide, against 93.4%. Passing each [Clairic]() span to [Jev]() to decide hide or keep raised precision to 91.1% while keeping [Clairic]()'s recall (96% on applicant names and case IDs, where Privacy Filter found 77% and 2.2%). It cost the most: $0.86 per 1,000 judgments against $0.26 and $0.17. Legal documents only. [Arjun Kumar, Clavia · Detecting private data using extraction + decision models · 31 Oct 2026]() ### Benchmark · arXiv 13 decision models against 15 generative LLM setups on 1,273 four-option MedQA questions. Decision models scored 67.71% to 89.08%, the generative setups 78.32% to 96.86%. [Mercury Decide]() was cheapest at $0.0069 per 1,000 questions and [Clef]() Omni fastest at a median 0.291 seconds. Answers given a probability of 0.90 or more were still wrong 1.71% to 12.77% of the time, depending on the model. [Freyer et al. · Decision models vs generative LLMs on medical questions · arXiv 2610.38448 · 28 Oct 2026]() ### Bias audit · arXiv [Jev]() and five open decision models on gender-swapped yes/no decisions about people, BBQ and StereoSet. Five of the six gave women significantly more favourable answers than men (+0.05 to +0.10 in log-odds), and swapping gender changed 1 to 3.5% of top answers. The makers' recommended confidence gate removed many biased answers but raised the stereotyped share of those it kept on ambiguous BBQ gender questions. [Havinga · When the Top Answer Decides · arXiv 2610.39747 · 29 Oct 2026]() ### Field report · independent A team put [Jev]() into a multi-agent build pipeline for Claude Code for 16 days, about 5,300 calls, as advisory checks, an on-demand tool and binding gates. It was accurate where the majority answer was enough and mixed elsewhere; 172 of 240 scope judgments came back uncertain or with insufficient evidence. On the claim-support gate, a strong model solved the task from the same excerpt, so the gap was [Jev]()'s. They removed it and never measured the token savings they were after. The anonymised data is published, with a script that rechecks 103 of the reported figures. [sash-ua · Accurate and Useless · 30 Oct 2026]() ### Field report · maker Squigdle, a daily drawing game, sends every doodle to [Jev]() as a 512×512 image plus its title and acts on the probabilities with its own thresholds. After about 1,000 doodles the misses are the occasional false alarm, and moderating all of them cost 6 cents in total, about $0.00006 a doodle. These are one developer's figures from a live game. [James O'Halloran · Moderating 500 daily anonymous doodles with Jev · 30 Oct 2026]() ### Field guide · independent Across 20 jobs drawn from TypeSafe's 21 official cookbooks, plus image tests, [Jev]() got 155 of 163 core examples right and 132 of 144 edge cases built to push its known weaknesses. Its main failure: when the right answer isn't an option, it picks the nearest one, often confidently. A replay after the locked tests, adding only a "none of these" exit to the menus, raised correct handling from 11 of 14 requests to 13. 1,463 requests cost $0.0331. Small hand-written samples, not a benchmark. [WeBuildDatabases · Jev, job by job · 29 Oct 2026]() ### Moderation simulation · independent In a simulated moderation queue, a wave of privacy-invading posts arrived that the existing router did not prioritise. Claude drafted one yes/no question in 12 seconds from 20 of those posts and 20 ordinary ones, and a fine-tuned version of the open decision model [Laya]() asked it of every new post. At 100 extra posts an hour, the queue needed one extra moderator instead of three, the same staffing as an oracle rule; the routing threshold was set with labelled calibration data. In a separate test, base [Laya]() answering a fixed money-back question scored AUC 0.68 on consumer complaints, against 0.89 for the complaint's product field. Three seeds; staffing counted in whole moderators. [Enayat Hassani · Exposure-aware routing · 29 Oct 2026]() ### Safety guardrail · USC Put in front of a small LLM (Qwen2.5-1.5B) as an input gate, [Jev]() cut how many of 720 harmful-labelled prompts got an answer from 527 to 17, but refused 27 of 250 safe prompts, against 3 for the LLM alone. A tiered gate that also reviewed borderline answers let 40 through and refused 10 safe ones. Answered counts are a proxy for harm, not checked completions. [Cognetti, Krishnamachari (USC) · Jev as an LLM Safety Guardrail · 2026, code and tables]() ### Video questions · independent Asked about video, [Jev]() 1.14.0 answered 463 of 658 native video questions correctly (70.4%), but pinning down when an event happens was weak: 10.0% mean average precision with three ordered frames, against 0.3% for a basic CLIP baseline, at temporal IoU ≥ 0.50 on the study's mapped action subset. [Eventum · Jev Video Lab, first study · 28 Oct 2026]() ### Bias audit · independent One identity word moves [Jev]() far more than a name does. On jev-1.13.0, "You are a Republican. Do you want to sign a carbon-tax petition?" got P(yes) 0.15; the Democrat version 0.73. Explicit labels shifted decisions by 3 to 15 percentage points, while Black- vs White-associated names moved them 0.2 (95% CI -0.6 to +1.0). Offering an "equal" option shrank most race, gender, religion and orientation gaps, but large gaps remained for age, undocumented status, class and atheism. 134,772 answers across jev-1.13.0 and 1.14.0, for USD 0.32. [Isaac Bitran (Brown University) · Jev bias audit · Oct 2026]() ### Head-to-head · independent On facts about other companies, Parallel's web-searching [Decisions API]() beat [Jev]() answering from memory: 83 to 98% vs 43 to 70% balanced accuracy on four questions about 50 YC companies. Across that test and a 30-topic publishing test, [Jev]() put 116 of 230 answers between 0.4 and 0.6. On "have we already published this?", handing [Jev]() the 10 closest sitemap URLs took it from 67% to 97%, at about 3% of Parallel's price. One run, one day, small sample. [Vedant Bhayani · Jev vs Parallel on two of Eric Siu's Jev gates · 28 Oct 2026]() ### Hallucination judges · arXiv Five decision models judged 9,852 items for five kinds of hallucination against gpt-6-sol. The best matched it on context inconsistency and came within 3 points on factual errors, but trailed by 4 to 8 on logic and instruction-following. A confidence cascade of decision models matched or beat gpt-6-sol on three types at 1 to 3% of its cost. [Liu, Xue, Wang, Mu, Ma, Feng, Li · Decision Models as Judges for Hallucination Detection · arXiv 2610.37229 · 27 Oct 2026]() ### RAG conflicts · arXiv Asked why retrieved sources disagree, 10 of 11 decision models from 7 providers matched or beat Gemini 2.5 Flash (61.3%), at 4 to 39x lower billed cost (two were free) and 2.6 to 6x lower latency. Gemini 3.8 Flash (70.0%) was more accurate than 8 of them, but sending each model's least confident quarter to it was statistically indistinguishable from Gemini 3.8 Flash for 8 of 11, at 25 to 37% of its cost. [Khatri, Jain, Kapoor, Sah · Do We Need an LLM to Know Why Sources Conflict? · arXiv 2610.36915 · 27 Oct 2026]() ### Vendor comparison · Voyage AI When a collection already has labelled examples, learning from them beat [Jev](), which labels each document cold. Across 29 classification tasks, Voyage's experimental voyage-json embeddings plus the raw fields, fed to a ridge regression trained on the labelled documents, averaged 88.3 AUROC against [Jev]()'s 77.6; Voyage counts that as 48% fewer errors (error = 100 minus the score). Voyage's own test of its own models. [Voyage AI blog · How to create your own Jev using Voyage models · 27 Oct 2026]() ### Option-order test · arXiv Shuffling the order of the answer options across 1,700 questions barely moved [Jev]()'s accuracy (under two points), but the set of questions it ever got wrong grew by 27 to 29%, more than plain repetition caused. Across two [Jev]() versions, twelve [Jev]() reproductions and six generative LLMs, [Jev]() showed no general advantage in keeping correct answers. [Song, Zheng, Ma, Zhou · Understanding Jev-Style System One Models: Does Structure Reduce Selection Bias? · arXiv 2610.35826 · 26 Oct 2026]() ### Clinical trials test · arXiv Screening 2,048 patient-condition pairs for clinical trials, with questions batched per patient, [Jev]() matched Qwen Flash (thinking off) on two datasets (macro-F1 0.652 vs 0.660, 0.807 vs 0.814) and beat it on MIMIC-III (0.768 vs 0.534), mostly by flagging when a record lacked the information (56.3% recall vs 9.9%). In batch mode it was faster and an estimated 53 to 74% cheaper per condition; asked one condition at a time, the cost edge did not always hold. [He, Gao, Li, Zhang, Wei, Zhang, Yuan · Can JEV Support Clinical Trial Criteria Assessment? · arXiv 2610.35721 · 26 Oct 2026]() ### Retrieval test · arXiv DecisionRAG uses a decision model to browse a corpus from broad topics down to passages, with no first-stage retriever and no task-specific training. It beat dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions, with median query latency of 0.68 to 1.76 seconds across four benchmarks. [Wan, Jünger, Du, Liu, Krishnan, Weinberger · Retriever-Free Retrieval: Navigating Documents with Decision Models · arXiv 2610.35868 · 26 Oct 2026]() ### Reranking test · arXiv Across six Qwen backbones from 0.6B to 27B and ten TREC DL and BEIR collections, decision fine-tuning never made the better reranker. Up to 9B, plain reranker fine-tuning won by 5 to 29 points nDCG@10 and decision fine-tuning added nothing over asking the untuned model; at 27B the two tied. The authors' advice: if you need a reranker, fine-tune a reranker. [Dinzinger, Ghosh Dastidar, Granitzer · System One models Are Not Better Rerankers · arXiv 2610.32318 · 23 Oct 2026]() ### Email routing test · arXiv On an owner-labelled Italian and English inbox (231 emails in the main comparison), logistic regression trained on about 200 labels had the highest accuracy on every field, though statistically tied with Gemma 4 on message type and with hosted [Nimble]() on urgency, ahead of local decision models ([Tev1](), [Nimble]()). When used to choose actions, it ignored emails that needed action more often than hosted [Nimble]() or Gemma 4 (18% against 4% for the best LLM), until a cost-chosen threshold halved that. [Federico Lucca · Measure Before You Train · arXiv 2610.32955 · 23 Oct 2026]() ### Pathology test · arXiv PathJev-P, a decision head on MedGemma-1.5-4B, beat a generative model trained on the same backbone and data, the pathology model CONCH and a general medical decision model on balanced accuracy (0.541 against 0.492, 0.345 and 0.213). Its calibration edge depended on how the generative model's probabilities were read, and a simple linear probe on a pathology foundation model was better calibrated in every setting. [Tommasino, Rinaldi · PathJev · arXiv 2610.32478 · 23 Oct 2026]() ### Search judging · independent Grading 200 held-out Amazon shopping-search results, [Jev]() 1.14 agreed exactly with the human graders 55% of the time, against 43.5% for the best LLM tested (Gemma 3 27B), at a median 118 ms against the LLMs' 435 to 656 ms. Asked to rank eight search setups, [Jev]() ordered 26 of the 28 pairs the way the human labels did. [rickcrawford · jev-relevancy, on Amazon's ESCI shopping queries · Oct 2026]() ### Calibration test · independent Logistic regression with temperature scaling, rebuilt behind [Jev]()'s interface, matched [Jev]()'s accuracy on a 7-intent banking slice (0.975 each; [Jev]()'s calibration error was lower, 0.014 against 0.022, but the gap was within noise) and beat it on all 77 intents (0.888 against 0.859 at best). [Jev]()'s choice answers were well calibrated; its yes/no answers were underconfident on true cases. [ankur287 · jev-open-calibration, code and data for a Towards Data Science article · Oct 2026]() ### Language test · arXiv Four decision models on 13 Vietnamese classification settings. With English label descriptions, [Jev]() ranked first by macro-F1 in 11 of them. The weak spot was diacritics: strip the tone and vowel marks and each model's mean macro-F1 fell 9.5 points ([Jev](), the least) to 15.5 (Quyet), against 0.5 to 1.1 for slang and 2.9 to 4.5 for typos. [Ha, Phuc, Quy, Tran · How Well Do Decision Models Transfer to Vietnamese? · arXiv 2610.31422 · 22 Oct 2026]() ### Legal test · arXiv Swap the correct case holding out of the options and see who admits none fit. Six generative LLMs still picked one in 38-86% of those items; the most cautious of four decision models did so in 30%. On the standard test, a decision model matched the best LLM (88.1% vs 88.2%) at 1/55 of the cost, and accepting the LLM's answer only when a decision model agreed halved its wrong citations. [Demir · Commit or Abstain? · arXiv 2610.29961 · 21 Oct 2026]() ### Clinical judging · arXiv As a judge of 118 paediatric clinical AI answers, [Jev]()'s agreement with physicians matched or beat seven of 12 LLM judges, including Claude Opus 5.5 and GPT-6. Frontier judges cost 73 to 294 times more and took 7 to 57 times longer. [Gershon, Cohen, Ariel, Bilavsky-Yarden · Decision language models as alternatives to LLM judges · arXiv 2610.28826 · 20 Oct 2026]() ### Independent index Artificial Analysis launched a Decision Index covering knowledge, reasoning, tool calling and safety classification. At launch [Jev]() 1.13 led the decision models with 58, ahead of [GLiDE]() and [Microsoft-Decision-1]() on 54. Its sharpest finding: decision models are overconfident. [Artificial Analysis Decision Index v1.0 · 13 Oct 2026]() ### Head-to-head On claim verification [Jev]() ties GPT-6-class models at 0.975, for a fraction of the cost. On Vals AI's LegalBench sample [Mercury Decide]() beats it, 0.869 to 0.740. Given a 1% error budget on claims, Mercury automated 94% at 0.8% error; [Jev]() automated 95% at 1.6%, over budget. [Vals AI · Jev evaluation 6 Oct 2026, Mercury Decide evaluation 8 Oct 2026]() ### Cost trick · arXiv Cut an LLM's reasoning short, then let [Jev]() pick the answer from the half-finished trace. Across 35 reasoner-benchmark pairs, [Jev]()'s mean accuracy beat forcing the LLM to answer at every budget tested, from 500 to 16,000 tokens; at 1,500 tokens it added 4.1 points. In offline replay on held-out questions, a tuned stopping policy was estimated to save 71.3% of the full run's cost for a 0.58-point accuracy loss. [Wang et al. · Truncated Reasoning Under a Budget · arXiv 2610.25768 · 19 Oct 2026]() ### Stress test · arXiv The study compared [Jev](), [Tev1-4B]() and Nimble-9B with two trained classifiers, with class frequencies shifted in simulation; the two open models were also run against their own Qwen backbones. Decision training was not uniformly better: [Nimble]() was better calibrated than its base on intent tasks but scored 12.0 Macro-F1 points lower on financial sentiment. The decision models were not uniquely fragile under shift, though single shifts could still push threshold risk up sharply; their main weakness was miscalibration they already had. [Iqbal, Nasim, Saleem · Typed Decision Models under Prevalence Shift · arXiv 2610.26325 · 19 Oct 2026]() ### Evidence audit · arXiv > In this early literature, the typed readout itself has not shown an independent accuracy advantage over comparable label-probability readouts. A review of the 28 papers posted in the nine days after [Jev]() launched on 15 Sep 2026. [Jev]()'s clearest gains were latency and cost, with accuracy gaps on harder tasks. The authors call it an early evidence map, not a verdict. [Typed Decision Models: An Early Evidence Audit · arXiv 2609.32160 · 26 Sep 2026]() ### Calibration study In SILT's preregistered test of 799 unpublished yes/no questions about coding-agent work and destructive shell commands, Perplexity's open [Decider]() V1.1-27B was the most accurate (94.7%) and the only one of seven models to pass SILT's chosen calibration rule. [Jev]()'s result was inconclusive. So were the results for OpenAI's [Decisions API]() and [Laya](). [SILT · Do Decision Models Know When They're Wrong? · 11 Oct 2026]() ### Security test Used as a prompt-injection detector, [Jev]() caught 82.90% of labelled attacks. Prepend a fake policy inside the untrusted text and it caught 3.88%. The authors separate this classifier bypass from a full agent compromise. [Datadog Security Labs · 15 Oct 2026]() ### Coherence test · arXiv Ask "is it X?" and "is it not X?" and the two probabilities should add to 1. [Jev]() misses by 0.064 on average. Reading the first token of an open 27B model misses by 0.293. Neither is fully coherent. [Li, He, Li · Beyond Calibration · arXiv 2609.33209 · 27 Sep 2026]() ### Independent leaderboard [JevBench]() v1.8.6, published 01 Nov 2026, measures 337 decision models and ranks 217 systems on its open-weights board. Capability averages Intelligence and Calibration within cost and median-latency caps. [Mammoth-1]() and [KnowLine]() 31B-Gen1 each score 85.2; the board marks their open-weight status Unknown. [Jev]() 1.13.0 is an unranked reference at 77.1. [SemIf](), formerly [OpenJev](), is #143 at 55.4 (Intelligence 27.1, Calibration 83.6). Hosted APIs have a separate board. [JevBench v1.8.6 · Benchmark Heaven · 01 Nov 2026 · not affiliated with TypeSafe]() ### First hands-on review Useful for ranking and classification, but it can carry bias you can't see: asked whether Bay Area cities were good, [Jev]() put Cupertino top and East Palo Alto bottom. With only a number coming back, that's harder to spot. [Simon Willison · 21 Sep 2026]() ### Open recipe · arXiv Without any decision fine-tuning, the instruction-tuned Gemma 4 31B already scores 79.6 on the authors' new OpenDecisions set, against [Jev]()'s 80.1. The authors' published recipe makes an open [Jev]() possible: after following it, the model scores 84.4 (the authors' own numbers). [Geuter, Cheng, Alvarez-Melis · Releasing the Inner Jev · arXiv 2610.17205 · 12 Oct 2026]() ### Replication · arXiv Political-science coding tasks from published studies, re-run on [Jev](), GPT-6 Luna and an open Qwen3.8-27B. [Jev]() matched or came close to both, but had no cost edge over OpenAI's batch prices, and its probabilities were better calibrated than Luna's but not consistently better than Qwen's. The authors' verdict: unless you need speed, the main advantage is ease of parsing. [Denney, DiGiuseppe · JEV versus LLMs · arXiv 2610.06625 · 6 Oct 2026]() ## We asked one model the same thing three ways and got different answers Our benchmark · v0.1 [Code and raw data]() | Path | Median latency | Output | | --- | --- | --- | | Generated JSON | 97.7 ms | 11 output tokens | | Schema-constrained | 74.4 ms | 7 output tokens | | Direct readout | 24.6 ms | 1 output token | - constrained vs direct: 76% - generated vs constrained: 68% - generated vs direct: 48% **The fastest path was also the least varied.** On this 0.6B model the direct readout answered *technical* in 69 of 75 calls and never once said *billing*, on a set that includes "my card was charged twice". Without labels we can't score accuracy, but an answer that barely changes is a warning sign. Whether a purpose-trained decision model avoids this is the open question. - Model: Qwen3 0.6B, Q4_K_M - Runtime: Ollama 0.33.2 - Machine: Apple M3 Pro, 36 GB - Calls: 25 inputs × 3 repeats × 3 paths - Temperature: 0 - Run: 17 Sep 2026 Medians over 75 timed calls per path. This measures mechanics, not accuracy: there are no ground-truth labels, so "disagree" means the paths differ, not that one is right. Local latency says nothing about any hosted API. [Raw results](https://decisionmodels.ai/bench/output-paths-v0.1.jsonl) · [Summary](https://decisionmodels.ai/bench/output-paths-v0.1-summary.json) ## Earlier research that Jev builds on When [Jev](https://decisionmodels.ai/models/jev) launched on 15 September 2026, [a GLiNER2 co-author](https://x.com/george_onx/status/2100293114808119379) answered by pointing readers to weights already on Hugging Face. The papers trace which parts of the idea predate [Jev](https://decisionmodels.ai/models/jev) and which are new. Every entry is checked against arXiv or Hugging Face. ### 2017: On Calibration of Modern Neural Networks Guo, Pleiss, Sun, Weinberger Showed that modern networks are systematically overconfident, and that a probability is not a calibrated probability until someone measures it. [Source]() ### 2018: BERT Devlin, Chang, Lee, Toutanova Sequence-classification heads over a bidirectional encoder: state in, bounded label out, no text generated. [Source]() ### 2019: Zero-shot classification via entailment Yin, Hay, Roth Arbitrary label sets declared at inference time with no per-task training. The 'declare your options' interface, seven years before [Jev](). [Source]() ### 2023: Function calling, then Structured Outputs OpenAI Schema-valid objects straight from an API. Structured Outputs added strict JSON Schema enforcement in August 2024. [Source]() ### 2023: Guided generation Willard & Louf Finite-state-machine constrained decoding: the answer space enforced during generation rather than validated after it. [Source]() ### 2023: GLiNER Zaratiana, Tomeh, Holat, Charnois A compact open-weight encoder that matches any entity type you name at inference time, in parallel rather than by generating. Not a general decision model, but the closest open ancestor the community pointed to. [Source]() ### 2023: Llama Guard Inan et al., Meta A safety classifier with bounded labels as the whole product, built as a language model that outputs its verdict. ShieldGemma followed in July 2024. [Source]() ### 2025: GLiNER2 Zaratiana, Pasternak, Boyd, Hurn-Maloney, Lewis (Fastino) Multi-task extraction and classification through a schema-based interface, several tasks in one forward pass, running on CPU. [Source]() ### 2026: GLiGuard Fastino A 0.3B schema-conditioned encoder judging prompt safety, response safety, refusal, 14 harm categories and 11 jailbreak strategies in one non-autoregressive forward pass. Apache-2.0 weights on Hugging Face. [Source]() ### 2026: Jev TypeSafe AI A hosted model trained specifically for calibrated typed decisions, with Choice, Score and Noul primitives. The weights stay locked: as of 20 Oct 2026, TypeSafe's Hugging Face org has published eval datasets but not a single model. [Jev]() [Source]() ## Where to read more - Start here: [Introducing System One Models and Jev]() · [Source]() TypeSafe AI's launch post, 15 Sep 2026, gave the category its name. - Start here: [Jev]() introduces a new shape of LLM · [Source]() Simon Willison's hands-on write-up, 21 Sep 2026. The clearest independent explainer. - Start here: A Deep Dive into Decision Models · [Source]() Joe Barrow, an NLP researcher at Adobe Research, explains how decision models work through an open one, Perplexity's [Decider]() v1.2: the architecture, how it is served, what its design says about training, and when to use one. An X article, 30 Oct 2026. - Start here: Latent Space: System One models for Prod, not God · [Source]() Interview with Diogo Almeida, TypeSafe CEO, 21 Sep 2026. The fullest statement of their thesis, including RLCD. - Start here: The launch thread on Hacker News · [Source]() Nearly 2,000 points and over 500 comments. The sharpest pushback is in here. - Papers: Typed Decision Models: An Early Evidence Audit · [Source]() Reviews the first 28 papers and proposes a 14-item evaluation checklist. 26 Sep 2026. - Papers: Beyond Calibration: do typed-decision probabilities obey the axioms? · [Source]() A label-free coherence test, run on [Jev]() and an open 27B model. 27 Sep 2026. - Papers: Decision models as layers in a pentest agent · [Source]() An exploratory case study of [Jev]() inside an LLM pentest agent, plus a discussion of [Laya]()'s published specs. 24 Sep 2026. - Benchmarks: [JevBench]() · [Source]() Independent leaderboard on intelligence, calibration, speed and cost. Not affiliated with TypeSafe. - Benchmarks: DecisionBench · [Source]() Ours. Three output paths on one model, with the hardware published. - Run one yourself: [Ollama v0.35: decision models]() · [Source]() Native /v1/systemone support on your own machine, from 28 Sep 2026. - Run one yourself: [SemIf (formerly OpenJev)]() · [Source]() Typed option probabilities straight from an open model, in your browser. No waitlist. - Run one yourself: [bnsd55/jevmlx]() · [Source]() Jev-style parallel constrained decisions for any MLX model on Apple Silicon; typed schema-valid JSON in one forward pass. - Run one yourself: [kw2828/OpenJev]() · [Source]() Browser decision playground and reproducible experiments on memory, uncertainty and control. - Community: AbdelStark/awesome-typesafe · [Source]() Community-maintained list of TypeSafe, System One and [Jev]() resources. Scoped to that ecosystem, not a general decision-model index. - Community: System One Models · [Source]() Another independent directory of decision models, with recipes and guides. - Community: @decision\_models on X · [Source]() Where this site posts new releases first. ## Get an email when a new model ships You'll also hear about new benchmarks and notable projects. You confirm your address before we send anything. We store your email, when you signed up and confirmed, where you signed up from, and a one-time confirmation token. We store nothing about your device or location.