Everything about TypeSafe AI’s Jev, in one place
Jev is a model that answers typed questions about your program state instead of writing text back. This directory collects what exists around it: where the documentation is, which SDKs and gateways reach it, which open implementations speak the same protocol, and who has measured it.
Start here
See the full directory- Guides & coverage
What Is Jev? (Firecrawl)
The best single starting point among the third-party explainers. It gets past the headline numbers into six concrete jobs a decision model is actually good at, including reranking retrieval results, checking an agent's citations against the page they came from, and screening fetched pages before they reach a model that will act on them.
Firecrawl2026-09-23 - Model & officialOfficial
Jev
The model this site is about. Jev takes your program state plus a set of typed questions and answers all of them in one pass, returning a choice, a score or a probability rather than a sentence. Because the answer is picked from a schema you wrote, it cannot come back off-schema. TypeSafe describes it as the first of a class they call System One models, and publishes speed and price comparisons against chat models; those figures are the vendor's own.
TypeSafe AI2026-09 - Evaluations
Jev out-of-distribution calibration test
Asks the question that matters for a model sold on calibration: does the calibration survive on a task the model cannot have seen? It pairs public benchmarks with rule-generated tickets whose labels never appear in the text, and reports error against a noise floor rather than in isolation. The useful conclusion is that miscalibration differs by question type, so one temperature for the whole model is the wrong unit.
scienthoon7 stars2026-09-19 - Open implementations
LocalJev
A bridge rather than a reimplementation: it keeps the request and response shapes, but gets its probabilities by asking an ordinary local chat model for a JSON number and then validating and normalising the result. That lets it run against a plain OpenAI-compatible server with no patched backend, and the README is explicit that a self-reported probability is not the same thing as a logit read. Check the calibration before you threshold on it.
GitHub Next776 stars2026-09-18 - Open implementations
OpenJev
An open decision server that speaks the same wire protocol on an open model, so the official SDKs can be pointed at it. Its distinguishing choice is to read answers out of the model's own token probabilities instead of parsing generated text, which is closer to what the hosted model does. The price is a patched inference backend and a GPU or Apple silicon machine to run it on.
razorback16427 stars2026-09 - Evaluations
PriorBench: pre-registered evaluation of Jev
The methodologically strictest evaluation listed here: fifty predictions were written down and timestamped before any data was collected, and the repository reports how many of them were falsified as prominently as how many held. The finding that travels furthest concerns the confidence value, which the authors found flat across most of its range and informative only at the very top.
PriorBench0 stars2026-09-20 - Model & officialOfficial
Introducing System One Models & Jev
The launch post, and the clearest statement of the design bet: give up string generation entirely, and train instead for probabilities that mean what they say. It names the training approach Reinforcement Learning for Calibrated Decisions, and frames the model as a function call rather than a conversation. Read it for the reasoning; read the evaluations elsewhere on this site for how it holds up.
TypeSafe AI2026-09 - DocumentationOfficial
TypeSafe AI documentation
Start here if you are writing code. The docs are organised around the three question primitives, then around what to do with the confidence value: threshold it, route on it, or decompose a hard question into several easy ones. There is an llms.txt index if you would rather hand the whole thing to an agent.
TypeSafe AI2026-09 - Gateways & access
Jev on Vercel AI Gateway
A route to the model that does not need a TypeSafe account of your own: the gateway exposes Jev under its own model id, and the AI SDK calls it through a dedicated evaluate function rather than a chat completion. The gateway lists the model id typesafe-ai/jev and the same published input price as TypeSafe. Worth knowing about, because one of the independent evaluations listed here — the out-of-distribution calibration test — ran its calls through it.
Vercel2026-09-16
Browse by category
- Model & official3
Jev itself: the model card, the launch announcement and what TypeSafe AI publishes about it.
- Documentation4
The official docs: quickstart, the question primitives, patterns, cookbooks and the HTTP API reference.
- SDKs & clients2
Client libraries that wrap the decision endpoint so questions stay typed in your own language.
- Gateways & access2
Platforms that relay Jev, so you can reach it with a key you already have.
- Open implementations2
Servers that speak the same wire protocol on open models, locally or on your own hardware.
- Integrations & tools2
Agent frameworks and developer tools that call a decision model in the loop.
- Guides & coverage4
Third-party explainers, walkthroughs and news written after the September 2026 launch.
- Evaluations4
Measurements with a published method you could rerun, rather than a headline number.
Recently published
- Guides & coverage
What Is Jev? (Firecrawl)
The best single starting point among the third-party explainers. It gets past the headline numbers into six concrete jobs a decision model is actually good at, including reranking retrieval results, checking an agent's citations against the page they came from, and screening fetched pages before they reach a model that will act on them.
Firecrawl2026-09-23 - Guides & coverage
A coding guide to Jev (MarkTechPost)
A hands-on walkthrough that builds up from a single question to the patterns you end up needing anyway: thresholds kept in code so risk tolerance stays reviewable, many questions batched into one call so the state is only sent once, and narrow atomic judgements combined with weights you control instead of one broad question.
MarkTechPost2026-09-23 - Evaluations
PriorBench: pre-registered evaluation of Jev
The methodologically strictest evaluation listed here: fifty predictions were written down and timestamped before any data was collected, and the repository reports how many of them were falsified as prominently as how many held. The finding that travels furthest concerns the confidence value, which the authors found flat across most of its range and informative only at the very top.
PriorBench0 stars2026-09-20 - Evaluations
Jev out-of-distribution calibration test
Asks the question that matters for a model sold on calibration: does the calibration survive on a task the model cannot have seen? It pairs public benchmarks with rule-generated tickets whose labels never appear in the text, and reports error against a noise floor rather than in isolation. The useful conclusion is that miscalibration differs by question type, so one temperature for the whole model is the wrong unit.
scienthoon7 stars2026-09-19 - Guides & coverage
TypeSafe AI releases Jev (MarkTechPost)
A short news write-up of the launch that is careful about whose claim is whose. It repeats the published price and latency figures, notes that the comparison demo is the vendor's own, and points out that a price cannot be shown to be unsubsidised from the outside. Good for a quick factual picture of what actually shipped.
MarkTechPost2026-09-19 - Open implementations
LocalJev
A bridge rather than a reimplementation: it keeps the request and response shapes, but gets its probabilities by asking an ordinary local chat model for a JSON number and then validating and normalising the result. That lets it run against a plain OpenAI-compatible server with no patched backend, and the README is explicit that a self-reported probability is not the same thing as a logit read. Check the calibration before you threshold on it.
GitHub Next776 stars2026-09-18
