LLMs write essays. Jev makes the call. This is the community's living index of what people are actually shipping with TypeSafe AI's System One decision model — real repos, real measurements, and the patterns that already work.
Jev is TypeSafe AI's first System One model — a decision engine that returns typed, calibrated judgments (Choice / Score / Noul) instead of free-form prose. It runs in 70–500 ms at ~$0.042 per million input tokens, and output tokens are free. In other words: fast enough to sit inside a game loop, cheap enough to run over millions of rows, and constrained enough that your code can branch on the answer without parsing JSON.
Every entry below links to a public repo or a documented result.
Within 72 hours of launch, the community shipped the kind of things that used to need a frontier LLM and a prayer:
- A browser agent that books a Zürich → London flight in 7.1 seconds for $0.0039
- Trading bots placing a buy/sell decision every ~300 ms
- Computer-use agents deciding for $0.0002 per step — vs Opus 5 at $0.032
- Agent guardrails, code reviewers, drones, games, and bulk classifiers
Most of it is still experimental. This index exists so you can find the projects worth reading — and steal the patterns that already work.
You send Jev a state (a string, JSON object, or list of text) plus typed questions. It returns typed answers with probabilities — no invented fields, no malformed JSON, no off-schema values. Your code owns the control flow, the thresholds, and the side effects.
| Primitive | You declare | Jev returns |
|---|---|---|
| Choice | A list of options (up to 255) | The chosen option + probability per option + confidence |
| Score | An ordered rubric (2–10 levels) | A probability-weighted score + confidence |
| Noul | A yes/no statement | A single probability from 0 to 1 |
Three primitives is the whole API — and that's the point. Questions in one request are evaluated in parallel and in isolation against the same state, so asking ten questions costs roughly the time of one. That single property unlocks speculative fan-out, confidence-gated routing, and every pattern on this list.
The economics (vendor-reported): $0.042 / million input tokens, output free, 70–500 ms end-to-end. On TypeSafe's own four-workflow eval, Jev lands at 67.8% agreement with a frontier-model consensus at $0.0004 per case — essentially tied with GPT-5.6 Terra (67.9%) at ~1/76th the cost. Treat the multipliers as directional, not gospel: the honest independent numbers are lower.
| Project | What it does | Headline result |
|---|---|---|
| browser-use/jev-ultrafast | Browser agent with a dynamic indexed action space; a small LLM writes text only when a step needs it | Zürich → London Google Flights in 7.1 s / $0.0039 |
| jarrodwatts/jev-trader | Market maker on Monad/Kuru — one buy/sell decision per block | 81 ms model latency per block |
| awlevin/typesafe-computer-use | macOS computer use via OCR + Jev action choice | $0.0002 / decision vs Opus 5 at $0.032 |
| DevMortimer/pi-warden | Agent guardrails that steer instead of interrupt | 6 rule breaks → 0 across 150 paired runs |
| AnshChoudhary/typesafe-ai-firewall | Pre-execution firewall for agent tool calls — one Noul per hazard | 0% hard negatives blocked vs 39.2% with a single "is this dangerous?" prompt |
| Gaurav-Gosain/jev-sec-bench | Blind prompt-injection & vulnerable-code benchmarks | 96.5% injection accuracy, ECE 0.0588 |
| TokenTrim/jev-agent-failure-benchmark | Can a decision model diagnose what broke an agent? | Beat GPT-5.4 on every axis across 6,257 traces for $1.28 |
| realZachi/pg-jev | Natural-language WHERE clauses for PostgreSQL — no embeddings, no vector column |
129 rows in ~1 s for ~$0.0009 |
| tamaratran/fast-jev-compaction | Claude Code compaction → Jev keep/delete | Content stays verbatim |
| TheoLeeCJ/openjev | Open re-implementation on a frozen 4B model by reading option logits | Runs in the browser, no waitlist |
| fhshaik/typesafe-mario | Super Mario from emulator RAM as object-centric JSON | No screenshots sent to the model |
| devagrawal09/jev-review | Staged code reviewer + local dashboard | Risk matrix → evidence → severity → routing |
| thruwire/foreman | Supervisor over Codex workers | Architecture experiment for software factories |
| RomanSlack/jev-drone | Camera-only quadrotor in MuJoCo | Advisory at ~2.5 Hz; safety stays in code |
Evidence note: every headline result above is self-reported by the project author — launch-week artifacts, not independent audits. Treat them as proof-of-concept signal, not production case studies.
The fuzzy middle between brittle rules and expensive chat models — classify, route, score, and branch inside ordinary software.
- Support triage: department, urgency, frustration, refund intent, policy fit
- Model & skill routing in agent harnesses — send simple work to the fast tier, hard work to the strong tier
- Email fraud detection, résumé–job scoring, lead/ICP qualification
- Natural-language Postgres
WHEREclauses → realZachi/pg-jev - Semantic spreadsheet formatting, Home Assistant automations → AboveColin/HA-Jev
Cheap judgment over giant corpora — the economics that make "run it on everything" rational.
- 1,018 papers → 24 topics for ~$0.08 (median ~256 ms)
- 98,000 listing classifications in ten minutes; YouTube comments, reviews, ad labeling, CMS tagging, log clustering
- 50 M-row scoring enters the ~$20 range
- Semantic features for classical ML — e.g. 2,000 wine notes → CatBoost numeric features at 1.77 RMSE
Action selection at game, UI, and market clock rates — perception and safety stay in code, Jev picks the move.
- Official: Doom (~10 Hz, ~$7/hr), Wikiracing over high-cardinality link frontiers
- Community: Subway Surfers ×50 parallel, StarCraft (phyous/tsai-sc), Snake, Tetris, Pac-Man, 2048, stealth games
- Live conversation-state sensors; trolley-problem and moral-dilemma loops
Screens become actions — Jev chooses the operation, a small writing model fills in only the free text.
- browser-use/jev-ultrafast (flagship), plus Ying-Kai-Liao/jev-browser, jkudish, tontoko, vlad-terin
- Android UI agent: Uber route demo in ~21 s / 9 actions → droidrun/mobile-jev
- macOS computer-use via OCR → awlevin/typesafe-computer-use
- Voice → action: spoken "go back" mapped to a click in ~300 ms
Make the agent loop cheaper, safer, and composable — Jev as the load balancer above models, tools, and humans.
- Tool-call firewalls → AnshChoudhary/typesafe-ai-firewall
- Steering agents instead of interrupting → DevMortimer/pi-warden, AbdelStark/bicameral
- Compaction, claim verification, prompt-injection screening, agent-failure diagnosis
- MCP servers → jkudish/jev-mcp, itsmostafa/typesafe-mcp
Relevance without embeddings — score each candidate with a Noul and sort by the probability.
- Rerank benchmarks vs Cohere / ZeroEntropy → anessbelbati/jev-rerank-bench (nDCG@10 0.692 vs 0.691 — a tie)
- Citation grounding: does the quoted context support, contradict, or stay silent on the claim?
- Candidate selection for context windows; knowledge-graph entity alignment
Score, judge, and gate prompts, traces, tool calls, and claims — at a fraction of the LLM call you're protecting.
- Prompt-injection & vulnerable-code benchmarks → Gaurav-Gosain/jev-sec-bench
- Moderation pipelines — notably strong on non-English (Japanese: 36 misses vs OpenAI's 292 on 826 harmful texts)
- Pre-execution shell / tool safety checks; jailbreak screening on inbound and outbound messages
Reproducing the interface, not the weights — proof that the decision-layer idea travels.
- TheoLeeCJ/openjev and friends (logit reading, constrained parallel decoding)
- Local / browser / Apple Silicon / vLLM / SGLang experiments; calibration research and playgrounds
- Independent evals worth reading: Every's 777 judgments in <0.7 s (~25× faster than Claude, 6-of-7 defects caught); Near Here's 96% listing moderation; Archer Hume's 84.6% MMLU-Pro probe (third-party, not TypeSafe)
- TypeSafe AI · Introducing System One Models & Jev · Docs · Console
- Official SDKs: typesafe-sdk-python · typesafe-sdk-js · system-one-adapter-python (benchmark LLMs on the same interface)
- Bypass the waitlist: the Vercel AI Gateway route (
typesafe-ai/jev) - Community clients: Elixir, Ruby, Rust, .NET, PHP/Laravel, Go, Scala/ZIO — see the ecosystem on GitHub
This is a community index. PRs are welcome for:
- New measured use cases (include cost / latency / accuracy when possible)
- Missing high-signal repos — link the public source, not a description
- Corrections or better categorization
- Patterns that have proven reliable
Keep entries factual and link to source. Separate TypeSafe-reported numbers from your own measurements.
MIT — for the index itself. Individual projects retain their own licenses.
Last updated: 18 September 2026
Jev is only a few days old. Expect this list to grow fast!!!