Use it as a smart if-statement. Your code keeps the flow, the arithmetic and the final decision; Jev supplies semantic judgments as typed data that your code can act on.
One request can carry dozens of questions, evaluated in parallel against the same state. You pay for the state and question tokens. The training method is called RLCD (Reinforcement Learning for Calibrated Decisions): on a yes/no question, 0.5 means the evidence does not favor either answer; it is a reason to abstain, not a medium rating.
Jev 1.14 was released on 14 October 2026. Image input is experimental on 1.13 and 1.14; Box, which returns the location of a target in an image, requires 1.14. The examples and playground on jevbooks currently demonstrate text with the three decision types below.
Ready to try it? Sign up, get a key and send your first request →
Already using a language model as a judge? Read Jev vs an LLM judge →
Three common decision primitives
Every question you send has a type. All three examples below run against the same state — the toy on our home page:
{
"task": "Book a table for two at 7pm tonight.",
"observation": "Reservation confirmed — 2 guests, 7:00 PM, Table 12."
}noulA yes/no question. Returns one number between 0 and 1: the probability of yes. An absolute judgment, and the only type with no confidence field.
Does
observationconfirm that the goal intaskwas achieved?0.98 — yes. Change the observation to “Almost there…” and it falls to about 0.03.
choiceOne of up to 255 options. Returns the picked option, a probability for every option, and a confidence. It is a relative judgment, so always leave an `other` escape hatch — otherwise the model picks the tallest dwarf.
What happened on this step?
goal_met·subgoal_met·blocked·otherblocked 0.88 (goal_met 0.00, subgoal_met 0.12, other 0.00), confidence 0.84.
scoreA graded level. You write two or more levels, each describing a concrete situation rather than an adjective; Jev returns the probability-weighted mean, so the score can land between two levels.
How far did the booking get? 0 nothing done · 1 started · 2 one subgoal confirmed · 3 all but the final step · 4 fully confirmed
1.97 — between “started” and “one subgoal confirmed”.
Question keys are never sent to the model, so the whole meaning has to live in the instructions. Point at state fields with backtick paths like observation. The experimental box type in 1.14 returns four normalized edges and a confidence. A missing target still produces a box: pair it with a Noul question asking whether the target exists.
What to keep in code, what to give Jev
This split is the biggest lesson from the projects that use Jev best: everything countable stays on your side of the line.
| Keep in code | Give to Jev |
|---|---|
| Enumerating candidates — DOM nodes, dropdown options, diff hunks, file ids. The model answers “which one”, never a selector or a string. | “Which of these candidates fits best?” |
| Thresholds, budgets, retries, stop conditions, dedup, loop detection. | “Is the goal met?” “Is it stuck?” — as probabilities. |
| Counting, measuring, comparing dates. A rule that needs a number is not a question. | Semantic existence: “does this text contain X?” |
| Combining probabilities: min / max / weighted sums / tiered bands. | One atomic probability per question. |
The final correctness check — read the URL, the form value, the exit code. A done signal is evidence, not proof. | The option probabilities behind done and blocked. |
What Jev cannot do
These engineering limits come from TypeSafe’s published 1.13 jaggedness notes and our September experiments. Recheck behavior on your own examples when changing model versions.
- It reads literally. Ambiguous wording gives an ambiguous number.
- It cannot count and cannot do arithmetic.
- Date comparison is unreliable.
- Multi-hop reasoning is weak — chain the hops yourself, one question per hop.
- Irrelevant state rots the context: send only what this judgment needs.
- It has no defence against prompt injection carried inside the state.
- Noul and Choice answers are not arithmetically consistent with each other.
- It cannot generate: no text, no code, no summary, no explanation of its own answer.
A September 2026 text-checker experiment sent encoded image data as ordinary text and received HTTP 200 with answers near 0.5. That was not an image request. Images now go through assets with TypeSafe-Danger-Zone-Images: true; a probability near 0.5 is not a general detector for unreadable input. Validate what your tool accepts before sending it.
Price and latency
- Input price
- $0.042per 1M input tokens
- Output price
- freethere is no generated output
- Latency
- ~300 msp50 375 ms from our machine
Latency was measured on 2026-09-21 for text requests to api.typesafe.ai, with jev-latest resolving to jev-1.13.0. Prices and limits were checked in the official docs on 2026-10-22; these timings do not measure 1.14, images or long context.
| Jev | |
|---|---|
| Input price | $0.042 per 1M input tokens |
| Output price | free — there is no generated output |
| Latency | about 300 ms; p50 375 ms from our machine on the official API |
| More questions | same forward pass — you pay the question tokens, not another call |
| Default context | state plus all questions ≤ 64k tokens; state plus any one question ≤ 32k |
| Experimental long context | both limits rise to 262k with `TypeSafe-Danger-Zone-Long-Context: true`; slower and less accurate |
| Rate limits | set per organization; check console.typesafe.ai/rate-limits |
| Input | text and JSON; experimental images via `assets` and `TypeSafe-Danger-Zone-Images: true` |
Cost in practice: the 530 calls behind our largest experiment came to $0.027.
How jevbooks uses Jev
The gallery’s eight signal questions use the repository name and description, an existing summary and the first 700 README characters. A separate noul gate asks “is this actually about Jev?”; the signal questions ask what their supplied material supports: is Jev core to it, does it report measurements, does it describe how it handles uncertainty, is it runnable, does it show the System One pattern, how common is the problem it solves, is it a replica of the model, would you recommend it to someone learning Jev. These are model judgments about a limited text snapshot, not a full source audit or proof that a project runs.
A language model writes the card prose; the verdicts and every probability on a card are Jev’s, and the card says which is which. That is the whole editorial policy.
2112 projects are listed, as of 4 Nov 2026.
Frequently asked questions
Is Jev an LLM?
Jev is a System One model designed to return typed decisions instead of generated prose. Depending on the question type, it returns a probability, a choice, a score or an experimental image box. You cannot chat with it or ask it to write an explanation of its answer.
Can Jev write code?
No. Jev does not write code or prose. Use a language model to write the code and the questions; use Jev to answer those questions repeatedly at its per-token input price.
How is it different from a classifier I could fine-tune?
You define classes in natural language at call time instead of fitting new model weights to those classes. Jev is trained for calibrated decisions, but that does not guarantee calibration on your domain: compare it with your classifier on labelled examples, including errors and abstentions. Changing a rule avoids a model training run, but still needs evaluation before deployment.
Does Jev support Chinese input?
Yes. A Chinese state works, and you can mix languages, keeping the state in Chinese and the question instructions in English. English is the primary training language, so evaluate Chinese questions on your own examples. The 1.14 announcement reports multilingual improvements; our older measurements do not establish the size of that improvement.
Can Jev read images?
Yes, experimentally on Jev 1.13 and 1.14. Send images through assets with the TypeSafe-Danger-Zone-Images: true header; putting base64 in an ordinary text state is not enough. Box location requires 1.14 and should be paired with a Noul check for target presence. The jevbooks playground currently accepts text; audio and video are not direct inputs.
Where do I get an API key?
From the TypeSafe console at console.typesafe.ai/keys; the reference is at docs.typesafe.ai. Requests go to POST https://api.typesafe.ai/v1/systemone, and the same wire format is available through OpenRouter and the Vercel AI SDK. A plain fetch is enough; official SDKs exist for Python (typesafe-sdk) and JavaScript (@typesafe-ai/sdk). Step by step: jevbooks.com/en/jev-quickstart/.
The 16 design patterns
Each one is a problem you will run into, the way out, and the projects that took it.
- Viewfinder · State minimalism
- Menu · Candidate enumeration
- Question bank · Rules as questions
- Ghostwriter · Decide/generate split
- Questionnaire · Flat fan-out
- Prefetch · Speculative heads
- Blind review · Cross-check
- Funnel · Cascade
- Traffic light · Tiered thresholds
- Grey zone · No-judgment band
- Thermostat · Hysteresis
- Scorecard · Composite scoring
- Inspection · Contract validation
- Fail open, fail closed · Failure-mode knob
- Flight recorder · Journal & replay
- Check-up · Measured calibration
Jev in your stack
Projects that put Jev next to a tool you already use.
- Claude Code
- oh-my-claudecode · fast-jev-compaction · agent-beacon
- Codex
- agent-beacon · memsearch · agentconnect
- MCP
- photoshop-mcp · jev-mcp · JevRouter
- Chrome extensions
- VideoAdGuard · jev-voice-browser · unclutter
- OpenCode
- agent-beacon · memsearch · foreman
Next
Ask Jev about your own problemBrowse the galleryOfficial docsGet an API key