What's new
Rows added to the catalogue and rows retired from it, week by week (weeks start on Monday, UTC). Also as an Atom feed. Feed
Week of 2026-10-26
Added
-
XERJ
Local-first search engine and MCP server in one Rust binary. Its optional `rerank` stage asks Jev one `noul` per top hit (does `document` answer the query in the state?) and re-sorts by the probabilities. The authors report FiQA nDCG@10 rising from 0.24 to 0.36, about $0.64 for 648 searches.written by a model
-
NeoHorse
TokenRhythm's open-weight model family for agent workflows. NeoHorse-Jev-4B, built on NeoHorse-1-4B, answers Choice, Noul and Score by prefill-only inference behind a /v1/systemone-shaped runtime for vLLM, SGLang or Python. The authors report 77.70 across six text benchmark groups.written by a model
-
Strands Decider
Open decision model for Strands Agents workflows: a 2B Qwen3.5 model, plus Gemma 4 sizes up to 26B-A4B, served behind a /v1/systemone-shaped endpoint and answering noul, choice and score with a confidence. The authors report 180 of 231 public JevBench tasks for the 2B and 207 for the 26B-A4B.written by a model
-
狗头军师 Jev Chat
Chat-reply assistant (desktop overlay, Android floating assistant, iOS keyboard) built on Jev Android. From the visible chat or a screenshot it asks Jev the literal question, the other person's intent, a danger level, whether to reply now and the best action, then has Jev pick among drafted replies.written by a model
-
Phocinae-Largha-150M-v1
A 144.3M-parameter bilingual (English and Chinese) typed decision model fine-tuned from mmBERT-small, answering noul, choice and score with zero output tokens through a /v1/systemone-compatible server, and runnable on CPU. The card leads with 0.791 typed accuracy in English and 0.848 in Chinese.written by a model
-
jev4pg
Natural-language SQL and 41 semantic operators for PostgreSQL. Filtering, extraction, ranking, matching and verification go to Jev in bounded batches with validated answers; a choice option needs 0.55 and a noul 0.8 to count. The authors report 98% accuracy against 92% for pg-jev's request path.written by a model
-
jiwo
Open decision models on Qwen3.5 (jiwo-0.8b and jiwo-4b) that answer Jev's POST /v1/systemone request with a probability for every option in one forward pass, generating no text. The author reports public Decision Index 0.3 scores of 29.58 and 46.06 and a median request of about 50 ms on one H100.written by a model
-
jev-doc-search
Tutorial for finding the page that answers a question in a long PDF with Jev's `Choice`: one option per page for short documents, and for long ones a walk down a PageIndex tree, one small `Choice` per level over section titles and summaries, so neither the 255-option nor the 32k-token limit is hit.written by a model
-
Jevlint (codegirl-007)
Linter that checks code against plain-language rules: Tree-sitter extracts each code unit and Jev answers one `choice` per rule (pass, fail, optionally skip or abstain); a failure is reported only when its confidence reaches the rule's floor. Other backends: OpenAI, OpenRouter, Cloudflare.written by a model
-
Lichen
Drop-in, API-compatible replacement for Jev that runs open-weight models locally behind the same /v1/systemone endpoint, images allowed in the state. The author reports 204 to 207 of 231 public JevBench decisions with gemma-4-26B-A4B on a 24 GB laptop GPU, against 200 for Jev 1.13.written by a model
-
Bud Decision Studio
Desktop app for macOS, Windows and Linux that installs a local engine and runs eleven open decision models (Kev, Laya, Lev and others) behind TypeSafe's /v1/systemone API and the OpenRouter and Vercel gateway routes, with a playground, evaluation, decision history and a trainer.written by a model
-
deepseek-harness-jev (luobosibing2)
DeepSeek Harness plugin with twelve opt-in features using Jev as the judgment model: skill selection and glob-result ranking (one `noul` per candidate), completion and goal checks, drift reminders, interjection routing, tool-output and test-log filtering, and approval of single workspace operations.written by a model
-
JEV Trading Command Center
Paper-trading research app with a 3D command center. For a USDT futures symbol it asks Jev the direction for the next 1 to 4 bars (long, short or no trade), a `score` for entry quality and three `noul`s (valid setup, trap risk, acceptable risk); code gates every simulated entry on the answers.written by a model
-
jev-ui-test
UI test framework, a fork of jev-ultrafast: natural-language YAML cases run by pytest, Jev picks each step's action and target, and semantic assertions ask one `noul` per atomic claim about the page, with thresholds left in test code. The author reports a 458 ms median per action decision.written by a model
-
Jeva
Research System One model from the National University of Singapore (Jeva-4B). The encoder reads the state and each option separately, so option order does not matter, options can be cached and a null answer is explicit. Its Python `system_one` takes Jev-style Choice, Noul and Score questions.written by a model
-
jevcrawl
MCP server and CLI that crawls a website for an agent's question. Per page, one Jev request asks a `noul` for every text block (does it help answer the goal?) and two for every link (leads to the answer, or to a section about it); a last `noul` asks whether the kept excerpts fully answer the goal.written by a model
-
laya-php
PHP 8.4 and Laravel SDK for Laya, the open multilingual decision engine: describe a decision as a readonly class with enums, and laya-serve's /v1/systemone endpoint answers it as a typed object with a confidence per answer. It adds a fake for tests, caching, events and batch prediction.written by a model
-
LLM Debugger
VS Code extension that debugs by running the code. At each pause Jev picks the next action as a `choice` (step, continue, breakpoint, evaluate, finish) and answers two `noul`s: is the faulty line found, and is it this one. A generation model supplies only locations, expressions and the fix.written by a model
-
OpenJev (alanhuangyoo)
Open decision models (OpenJev-1.7B, 4B, 8B; package wev-ai) behind POST /v1/systemone, distilled from a browser agent's typed requests on live sites. Besides choice, yes/no and score, they pick a browser agent's next operation and element. The author reports 39 to 41 of 153 live tasks for the 4B.written by a model
-
StartLux-Decision
StartLux's open decision models (0.8B to 27B) with an engine and HTTP server answering choice, noul and score behind a /v1/systemone-compatible API on vLLM, llama.cpp or MLX. AutoThink reasons only below a probability gate. The authors report ranking first on Decision Index 0.3.2.written by a model
Week of 2026-10-19
Added
-
Yao Agents
Self-hosted agent workspace that runs agents across your own devices with a task board, isolated workspaces and an open API; it added Jev as a decision-model provider alongside its text models. The cited file tests Jev key validation against /v1/models.written by a model
-
jevgrep (dzhng)
CLI and agent skill that finds code by asking what it does: `jg "How are telemetry events recorded?" ./repo` has Jev judge relevance across folders, files and declarations and returns files and verbatim excerpts. The author reports the same 8 of 10 SWE-bench tasks solved at ~30% lower cost.written by a model
-
Jeff
An open 0.8B 'System 1' model with task adapters and GGUF builds, meant to sit in front of a large local model and make quick decisions; its game runner asks one choice per turn over legal moves via /v1/systemone. The author reports 94.6% mean accuracy over 8 adapters vs 86.6% for Qwen3.8-27B.written by a model
-
Ollaya
Runs open decision models locally the way Ollama runs LLMs: pull a model by name and serve it behind a /v1/systemone-compatible daemon, so the TypeSafe SDK works unchanged. The README reports 0.724 typed-decision accuracy for its recommended winnow:e4b against TypeSafe's 0.738 for Jev.written by a model
-
jev (jevos)
One CPU-only binary that serves an open model, jevos-v5, behind Jev's POST /v1/systemone: noul, choice and score answers with probabilities in about 30–130 ms on a laptop, no Python. Weights on Hugging Face and as GGUF; the author's 7-task benchmark has Jev ahead on accuracy.written by a model
-
vLLM Jev
vLLM-based serving engine for open Jev-shaped decision models such as Valen: scores choice, noul and score candidates through prefill with shared-prefix reuse, with image and video inputs and an MLX path for Apple Silicon. The author reports 4.3× over Valen's PyTorch reference on one video run.written by a model
-
Jeeves
PostHog's open 9B Jev-like model (Qwen3.5-9B, LoRA, pointer head) that answers noul, choice and score directly when confident and reasons first when not, with a diffusion drafter, training code and data. The authors report it ahead of Jev on JevBench's public tiers.written by a model
-
imajev
Small open models (2B, 4B, 9B) that read the photos, records and text a business already has and answer in the options you set, each with a probability and an explicit can't-tell, so a system acts when sure and hands the rest to a person. The author reports first place of 91 on JevBench v1.4.2.2.written by a model
-
beebots
Three trading bots race on OKX perpetual futures, paper trading by default: Jev (or Gemini) picks each move as a choice from the legal actions plus a score for conviction, with a deterministic fallback, a plain-code risk layer and a live dashboard.written by a model
-
10 Levels of Jev
Thirty uses of Jev arranged in ten levels, from one smart if statement in plain code to hooks and tools inside the pi coding agent and finally an agent that writes its own questions. A live lab, terminal runner, offline mock with 201 tests and a companion YouTube video.written by a model
-
OneJev
Multimodal open decision model on Qwen (0.8B, 4B, 9B and 27B) that returns calibrated probabilities for typed questions about screenshots, photos, video and text, with a local server the official TypeSafe SDK talks to unchanged. The authors report 69.49 for the 27B on the public Decision Index 0.3.written by a model
-
Building with TypeSafe Jev
Unofficial agent skill that teaches coding agents to design for Jev rather than treat it as a chat model: typed questions, thresholds on calibrated confidence, and links to 150+ community projects sorted by shape with a code sketch for each. Installs in Claude Code, Codex and others.written by a model
-
Jev-Code-Reviewer
Local reviewer for coding agents' pull requests: each change is classified P0, P1 or P2 by Jev choices against a written priority policy, and only P0 changes are shown by default. Posts nothing to GitHub.written by a model
-
Quicksilver
Claude Code skill that hands relevance calls to Jev: one noul per file, log line or ticket ('which of these files handle auth?'), asked in parallel and sorted by probability, so Claude gets a shortlist with confidences instead of reading everything.written by a model
-
Jev Plays Pokémon Red
Harness in which Jev beat Pokémon Red: it reads the emulator's memory, lists the legal options with facts (type matchups, damage estimates, distance to the objective) and Jev picks one. 16,150 decisions over 37 h 40 m for about $1.65, via the AI SDK on Vercel AI Gateway.written by a model
-
RSI-Jev
An agent loop that proposes hypotheses, pre-registers predictions, trains and retires open decision models answering noul, choice and score in the /v1/systemone shape; 496 experiments written up, failures included. The author reports 65.64 on the public Decision Index 0.3 set for the 27B VL model.written by a model
-
Jevry
Open-source desktop browser that works through websites from a plain-language task: a text model plans, Jev picks each action as a choice among what the browser can actually execute (never generated JavaScript or selectors), and Chromium acts with visible steps and stop controls.written by a model
-
jev-kit
Claude Code kit: a PreToolUse guard where a free pre-filter passes about 93% of tool calls and Jev judges the rest (allow, warn, rewrite or block; fails open), sub-agent sizing, and Jev wired into file search, browsing and review. The author reports 9/9 browser runs at 0.0008 USD of Claude spend.written by a model
-
JevEmbed
Toolkit that turns embedding models into choice, score and noul answerers: data synthesis, fine-tuning, a /v1/systemone server, a benchmark module and a playground. Released fine-tuned Qwen3-Embedding and KaLM checkpoints and a 1.67M-question dataset on Hugging Face.written by a model
-
Jevstiller
Proxy in front of a repeated Jev classification call: it forwards to Jev at first, trains a small local model on Jev's own answers and probabilities, and answers locally once it agrees with Jev on held-out traffic. The author reports 61.7% of Banking77 answered locally at 99.4% agreement.written by a model
-
grev
Unix filters that ask questions instead of matching patterns: `grev 'is a vegan meal' menu.txt` keeps the lines Jev says yes to, with related tools that use a choice's confidence or a path score. Pins jev-1.13.0 because thresholds are tuned per model version.written by a model
-
vibecheck
Python library that turns System One calls into plain values: check (yes/no), classify (one option), label (every option that applies) and score, with thresholds, an unsure answer and batching of several questions about the same data. Works against TypeSafe or any compatible server.written by a model
-
Flick
Local stdio MCP server that runs whole browser and macOS goals: each observe-decide-act step asks Jev a choice over actions keyed by observed element IDs, plus choices for which field and value to fill. Playwright browser driver and a Swift Accessibility driver.written by a model
-
JEV Safety Gateway
Reverse proxy between nginx and an OpenAI-compatible LLM backend that extracts each request's user input, asks Jev a noul on whether it is safe, and blocks harmful requests while passing the rest through; rotates over several API keys. Chinese-first docs.written by a model
-
Harness Router
Picks the next tool inside an existing agent loop: given a goal, observation and tool list, Jev returns a tool choice with confidence or a fallback to the host planner. Caches reusable choices, offers optional MCTS, and hooks into Codex and other harnesses or runs as an MCP server.written by a model
-
Jev-Style
Small local decision models on Qwen3.5 (2B and 0.8B) with a /v1/systemone server, an MCP server with decide, noul, choice and score tools, a Claude Code PreToolUse guard and an eval command for your own labels. The author reports 73.6% on JevBench v1.4.1's public items, self-run.written by a model
-
jev-lint
Fuzzy linter for coding agents: after an agent writes a file, each of the team's plain-language rules is asked as a Jev noul, and about 0.3 s later the agent hears which rules it broke. Can generate rules from the repository; Cloudflare and OpenAI judges as alternatives.written by a model
-
jev-opus
Runs Claude Code on Opus 5.5 with Jev re-picking the effort level when a prompt arrives and after every tool batch (up after a failing test, back down once it passes), sent as per-message effort statements so the prompt cache survives. Works as a local gateway or a headless driver.written by a model
-
DuoMind
OpenAI-compatible local server that pairs a small llama.cpp model with Jev: 18 decision points before, during and after generation (needs reasoning, safety, format, should stop, needs retry) steer the local model. Prompts go to TypeSafe; generation stays local.written by a model
-
JEV Sees
Gives Jev vision: local Florence-2 and CLIP models find the objects in an image, video or RGB-D frame and keep their IDs across frames, then Jev answers a choice, noul or score about each relevant object (bus colour, pedestrian risk, animal motion).written by a model
-
jev-enforce
Claude Code plugin that checks every reply and file edit against your CLAUDE.md or AGENTS.md: one Jev noul per rule in a single request, and broken rules are quoted back as a Stop or PostToolUse block so Claude fixes them in the same turn. Also a CLI for CI.written by a model
-
x402check
Pre-payment risk-check provider for x402 agents and wallets: sanctions lists, phishing feeds, transaction simulation and drainer-kit fingerprints, plus Jev nouls, a choice and a score that read what the agent acted on for injected instructions. Verdicts are signed attestations.written by a model
-
Decis
Self-hostable inference server for open decision models that implements the /v1/systemone contract, so the official TypeSafe SDK only needs base_url changed. One Docker image per engine with weights inside (laya-multilingual, kev-0.8b and two Jeff builds), plus a game playground.written by a model
-
Jev Browser Bridge
Runs a Jev browser agent on any CDP browser, even ones that never draw a page: it reads the DOM rather than positions and asks Jev one choice for the operation and one for its target. The author reports 59–65% on 172 MiniWoB++ episodes across six browsers, against 24–31% for a position reader.written by a model
-
typesafe-ai-php
PHP SDK for Jev: a TypeSafeClient, a fluent SystemOneRequest builder for state and typed questions, and a SystemOneResult keyed by question ID. PHP 8.2+, installed with Composer.written by a model
-
jev-browse
Helpers for browser-harness that let Claude Code or Codex hand off a whole website sub-task in one fast_run call, with each step decided by Jev and a typed reason returned when it must stop. A port of browser-use/jev-ultrafast; the author reports 1.6–3.0x faster and 2.1–5.3x cheaper runs.written by a model
-
typesafe-client (Go and Rust)
Unofficial Go (stdlib only) and Rust clients for the System One API with typed noul, choice and score questions, the official SDKs' retry policy, typed errors with request IDs, response validation, GET /v1/models and gateway support.written by a model
-
Jev Playwright
Playwright reporter that asks Jev which existing end-to-end tests a code change could affect and skips the clearly unrelated ones before any browser starts. Sends changed paths, patches and test bodies by default; its README warns it can miss affected tests.written by a model
-
JevT++
C++20 library that turns runtime text, JSON or objects into enums, booleans and scores through compile-time schemas, with full distributions and explicit abstention. Runs local Laya models in-process (ONNX Runtime or ggml) or Jev through an opt-in remote backend; latency metrics and a Mario demo.written by a model
-
solvi
Python runtime with a System One API adapter: Jev, an LLM or a local checkpoint proposes typed answers; hard checks, rules and calibrated thresholds decide, abstain or escalate to a person, and every decision is a hash-chained replayable trace.
-
Jev Civilizations
Small strategy game with a live demo: four tribes run their own lives while you only change the environment. Each turn Jev picks one action per tribe as a choice over the moves that are actually legal, a plain game engine works out the results, and an inspector shows what Jev was told.written by a model
-
1 Million Emojis
A shared 1,000 × 1,000 emoji canvas where Jev paints next to people: after each stroke one request asks a choice of emoji-and-placement options and a noul on whether the stroke looks unfinished, through Vercel AI Gateway with TypeSafe direct as fallback.written by a model
-
Jev Switch
Lightweight gateway that serves POST /v1/systemone and routes each call through an editable graph of upstreams (TypeSafe, Vercel, OpenRouter, Laya or a local server) with priorities and failover. Rust daemon with a React console, Tauri desktop app and Docker image. Chinese-first docs.written by a model
Retired
-
JevRev
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-curate
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-seo
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-superpowers
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-scout
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-git
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
rh-guard
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-13 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
zerosweep
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-09-29 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
datajev
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-09-29 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-screen-mcp
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-09-29 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
jev-skills
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-10-06 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
-
typesafe-oracles
Retired 2026-10-24: the repository has not resolved in any weekly refresh since 2026-09-29 (last on 2026-10-21) — deleted, renamed without a redirect, or made private. Kept here so the reference stays searchable.
Week of 2026-10-12
Added
-
VideoJev
Open System One video models that score temporal candidates and multiple-choice options for several queries over one shared video prefix, then return intervals, answers, and candidate probabilities without text generation.
-
JevGPT: Thinking Fast and Slow in Autonomous Research
Research agent from arXiv 2610.18272. GPT-5.5 plans searches and writes the answer; in BrowseComp-Plus QA, Jev answers a Choice over numbered source lines (where is the answer) and a Noul (does the region address it) per document region and query, in parallel. GPT gets the original lines.
-
Decision Models as Guardrails
Asks Jev 1.13.0, Jev 1.14.0 and 17 other systems, Bedrock Guardrails among them, the same 9,975 labelled guardrail checks in six suites. Jev answers each as a noul; 0.5 or more blocks, nothing tuned per model. Reports balanced accuracy with 95% intervals and USD per 1,000 checks at list price.
Week of 2026-10-05
Added
-
mu
Coding agent and desktop app built on pi that asks Jev at 38 decision points which tool-output chunks enter the context, which stale tool results to drop, whether a rule-flagged command was asked for and whether fetched pages or MCP output carry injected instructions.
-
Support-routing benchmark: 100 held-out cases
Compares Jev with Clef, Clef Flash, OpenAI's Decisions API and a GPT-6 Luna structured-output baseline on 100 held-out fictional support messages, each routed to one of five team labels; every case and recorded answer can be downloaded.
Week of 2026-09-28
Added
-
WaterSheep
Open, Apache-2.0 alternative to Jev: a ModernBERT fine-tune served locally on POST /v1/systemone that answers noul, choice and score questions, plus multi-label ones, with a probability for every option.
-
JevSpan
Zero-shot named entity recognition for Chinese and English: code enumerates candidate spans at punctuation, and Jev choice questions nominate a window per entity type, verify each nominee (type, none, mixed or partial) and pick its exact boundary, keeping each entity's probability.
-
jevbench
Preregistered, bias-corrected test of TypeSafe Jev's calibration under human disagreement (ChaosNLI, 100 labels per item)upstream description
-
jev-seo
Audits a business website for SEO and AI-answer-engine visibility. A deterministic crawler measures each page, then one request sends Jev a batch of narrow typed questions about it, and the answers are surfaced as probabilities in a dashboard.
Week of 2026-09-21
Retired
-
jev-mac-voice
Retired 2026-09-24: the repository returns 404 on both the API and the web while its owner's account still exists — deleted or made private. Kept here so the reference stays searchable.
-
jev-atlas
Retired 2026-09-24: the repository returns 404 on both the API and the web while its owner's account still exists — deleted or made private. Kept here so the reference stays searchable.