Awesome Jev Use Cases
September 18, 2026 · View on GitHub
Real use cases, working patterns, and hard-won guidance for building with Jev, TypeSafe AI's System One model — the model that returns typed, probabilistic decisions instead of text.
Jev is not a chatbot. You send it program state and typed questions; it returns typed answers with calibrated probabilities in 70–500 ms. This repository collects what people have actually built with it, and - more importantly - how to use it efficiently and effectively.
state + typed questions -> typed answers + probabilities -> your code decides
Status: research snapshot as of 2026-09-18. Jev is in early access and moving fast. Prices, limits, and aliases will change. Every claim below links to its source.
Start here
| If you want to… | Go to |
|---|---|
| Understand what Jev is and when to reach for it | docs/what-is-jev.md |
| Make your first call in five minutes | docs/getting-started.md |
| Learn the design method | docs/how-to-use-jev-effectively.md |
| Steal a proven architecture | docs/patterns.md |
| Avoid the mistakes everyone makes | docs/failure-modes.md |
| Decide Jev vs LLM vs a classifier | docs/model-selection.md |
| Verify it works on your data | docs/evaluating-jev.md |
| See the full use-case catalog | usecases/README.md |
| Browse real applications people built | reference/showcase.md |
| Copy a runnable example | examples/ |
| Find ready-made questions | reference/question-catalog.md |
The 60-second version
Jev is TypeSafe AI's first System One model, launched 15 September 2026 by Diogo Almeida (former OpenAI researcher, co-author of the InstructGPT paper) with roughly $40M in seed funding led by DCVC. The name comes from Daniel Kahneman's Thinking, Fast and Slow: LLMs are deliberate System 2 thinkers, and most decisions inside software only need a fast System 1 reflex.
You get exactly three question types, and that is the design, not a limitation:
| Primitive | You declare | Jev returns |
|---|---|---|
| Choice | A list of options (up to 255) | The chosen option, a probability per option, a confidence |
| Score | An ordered rubric (2–10 levels) | A score (can land between levels), a distribution, a confidence |
| Noul | A statement about the state (yes/no) | A single probability 0–1 |
All questions in a request are evaluated in parallel and in isolation against the same state. Adding a tenth question costs tokens but barely any time. That single property is what unlocks most of the patterns in this repo.
The economics, vendor-reported: $0.042 per million input tokens, output free, 70–500 ms end to end. On TypeSafe's own four-workflow evaluation Jev scores 67.8% agreement with a frontier-model consensus at $0.0004 per case, versus GPT-5.6 Terra at 67.9% and $0.0304. Read model-selection.md for the honest breakdown, including why those multipliers are softer than the headline.
What people have built
Jev launched on 15 September 2026. Within three days the community had shipped browser agents, trading bots, drones, code reviewers, and a semantic firewall. This is a sample - each entry links to the repository, and reference/showcase.md has the full detail plus 100+ more.
| Project | What it does | Headline result |
|---|---|---|
| browser-use/jev-ultrafast | Browser agent with a dynamic action space. Jev picks the operation and the target; a small LLM writes text only when the step needs it | Zürich → London on Google Flights in 7.1 s for $0.0039 |
| TheoLeeCJ/openjev | Reproduces the Jev interface on a frozen 4B model by reading option logits instead of generating | No waitlist; runs in a browser |
| tamaratran/fast-jev-compaction | Replaces Claude Code's lossy compaction summary with Jev keep/delete decisions | Kept content stays verbatim - nothing is rewritten |
| jarrodwatts/jev-trader | Market maker on Kuru MON-USDC. One buy/sell decision per Monad block | 81 ms model latency, 2 RPC round trips per block |
| fhshaik/typesafe-mario | Plays Super Mario Bros. from emulator RAM as object-centric JSON | No screenshots sent to the model |
| thruwire/foreman | Independent supervisor above Codex workers: is the work complete, tested, off-track, stuck | Framed honestly as an architecture experiment |
| devagrawal09/jev-review | Staged code reviewer: Noul risk matrix → file profiles → evidence selection → severity → routing | Ships a local dashboard |
| awlevin/typesafe-computer-use | macOS computer use via OCR + Jev action choice. A writing model only for free text | $0.0002 per decision vs Opus 5 at $0.032 |
| DevMortimer/pi-warden | Agent guardrails that steer instead of interrupting: rules, slop, stuck loops, done claims, irreversible calls | 6 rule breaks without it, 0 with it across 150 paired runs |
| AnshChoudhary/typesafe-ai-firewall | Pre-execution firewall for agent tool calls: one Noul per hazard, ALLOW/APPROVE/BLOCK decided in code | 0% of hard negatives blocked, vs 39.2% with one "is this dangerous?" question |
| realZachi/pg-jev | Natural-language WHERE clauses for PostgreSQL. No index, no embeddings, no vector column | 129 rows in ≈1 s for ≈$0.0009; re-thresholding is free |
| Gaurav-Gosain/jev-sec-bench | Blind security benchmarks on public corpora: prompt injection and vulnerable code | 96.5% injection accuracy, ECE 0.0588, 662 samples |
| kiarina/labs: safety judgment | Moderation vs the OpenAI Moderation API, plus shell-command safety checks before execution | Japanese moderation: 36 misses vs OpenAI's 292 of 826 harmful texts |
| anessbelbati/jev-rerank-bench | Jev vs Cohere Rerank 4 vs ZeroEntropy across 14 datasets, every raw response saved | nDCG@10 0.692 vs Cohere 0.691 - reported as a tie |
| RomanSlack/jev-drone | Camera-only quadrotor. Jev is advisory at 2.5 Hz; code owns safety at 50 Hz | ~110 calls per 65 s flight |
| TokenTrim/jev-agent-failure-benchmark | Can a decision model find what broke an agent as well as a frontier LLM? | Beat GPT-5.4 on every axis, 6,257 traces, $1.28 total |
| youkiti/tiab-review-plugin | Title and abstract screening for systematic reviews across six labeled medical datasets | 95.0% recall on 16,645 records at threshold 0.3 |
| zephel01/Jev-sample | The same decision asked as one 4-option Choice vs four precondition Nouls, with raw logs of 2,320 requests | 48.3% → 98.3% from question shape alone |
Browse by category: Agent supervision · Security · Code tools · Search · Classification · Computer use · Trading · Games and robotics · Domain apps · Open reproductions
No star counts here on purpose. Launch-week stars track attention, not quality - two of the most rigorous projects in this list sit at 1★ while a game demo sits at 200+. Sort by the measurement instead.
Evidence note. Every result above is self-reported by the project author and none has been independently reproduced. All of these are launch-week artifacts built in the first 72 hours. Where a project publishes raw outputs, showcase.md says so - those are the ones worth reading first.
Use-case catalog at a glance
Each category has its own page with the decision shape, evidence, code, and pitfalls.
| # | Use case | Decision shape | Reach for it when |
|---|---|---|---|
| 1 | Routing and triage | Choice + Score | One known category selects the next code path |
| 2 | LLM guardrails and verification | Noul + Score | You must check an LLM's input, output, or tool call |
| 3 | Agent harness engineering | Choice + Noul | A coding agent needs a fast semantic supervisor |
| 4 | Search, reranking, and RAG | Choice + Noul | Relevance decides what reaches the context window |
| 5 | Structured data extraction | Choice over candidates | Known fields must be recovered from messy text |
| 6 | Classification at scale | Choice | Millions of rows, bounded labels |
| 7 | Feature extraction for ML | Score + Noul | A classical model needs semantic signals |
| 8 | Real-time loops and games | Choice per tick | Decisions faster than human perception |
| 9 | Browser and computer use | Choice + Noul | A screen must become an action |
| 10 | Domain applications | All | Legal, finance, insurance, health, recruiting, commerce, moderation |
| 11 | Frontier and fun | All | The demos that show what the primitive can do |
Want the raw inventory of everything the community has shipped? See reference/showcase.md for the application showcase, or reference/ecosystem.md for the full census including SDKs and infrastructure.
The one-paragraph summary of how to use Jev well
Keep control flow in code and give the model narrow, atomic judgments. Decompose a broad question into several literal ones and recombine their probabilities with weights you own. Send only the state a question needs, because accuracy falls as irrelevant detail grows. Ask everything you might need in one call, then let code discard what was irrelevant. Gate actions on confidence, scaled to what being wrong costs. Do not ask Jev to count, do arithmetic, compare dates, or generate text. And test the thresholds on your own traffic before you trust them, because the calibration claim is asserted by the vendor and not yet independently reproduced.
The rest of this repo is the long version of that paragraph.
Ground rules for this list
- Evidence over vibes. Every project entry states what was measured and who measured it. Vendor-reported numbers are labeled as such.
- Use cases must use Jev. A generic classifier that merely resembles the pattern is not included.
- Launch-week artifacts are labeled. Many projects were built in the first 72 hours. Treat them as proofs of concept, not production case studies.
- Corrections welcome. See CONTRIBUTING.md. If a number here is wrong or stale, open a PR.
License
MIT. Not affiliated with or endorsed by TypeSafe AI.