Lesson 02 · Jev 101

How Jev works: state + typed questions

Forget prompt engineering for a moment. Every Jev call has the same shape: state in, questions in, typed answers out.

State (your context)

State is whatever the decision is about: a ticket, email, agent trace, document, game state, log blob—often unstructured text, sometimes structured program state (company emphasis). Official docs: a string, a JSON object, or an array of text. Text only — no image, audio, or video yet. Communityjev-ax-pilot therefore walks the Accessibility tree instead of uploading screenshots,MacWork does the same on a live Mac (14/29 is one run, not a success rate), andjev-tower lets code predict geometry — Jev is not visionPRESS.

ChatGPT parallel: the stuff you paste into the prompt.
Difference: you’re not also asking Jev to write the answer in English.

Official jaggedness note: accuracy falls when you stuff large irrelevant state (“context rot”). Filter. Send only what the decision needs. The reverse bug is just as real: if the fact is not in the state, Jev cannot invent it.jev-drone first saw Jev refuse climb every time — five range sectors had no vertical information, so “fly over it” was not inferable. Adding the obstruction’s top-edge height moved climb to p=0.93. That was a state-design bug, not a model failure. Jev is also not a vision model there: 15 Hz classical CV writes the JSON; Jev only answers at ~2.5 Hz; a 50 Hz reflex can override any judgment.NRS Navigator is the same lesson on a statute shortlist: Jev cannot recover a keyword the keyword stage omitted, and three nominated sources were never in the 30-candidate pool — Hit@5 12/18 vs 15/18 is that regression, not lawyer gold.PRESS

Jev also has no conversation history. Each request is self-contained — if last turn matters, your code must put it in this state.milanboers/jev-plays-pokemon does that on Pokémon Red: the harness re-injects recent dialog and actions every turn. Community demo, not TypeSafe.PRESS

Official System One page: like an LLM, it understands natural-language input; unlike an LLM, it returns typed decisions and probabilities rather than generated text. The name is Kahneman’s fast/intuitive System 1 as a metaphor for focused judgments — not a claim that Jev is human cognition.

Three question types

Multiple questions can be evaluated together in one request (company: independently, in parallel). Official theme: speculative fan-out — ask more than you might need, let code select. Official Choice page: up to 255 options. Communitywellposed restates that 255, plus official 64k / 32k, and shows a closed menu can answer at confidence 1.00.MacWork folds groups instead of ranking options away when the menu is larger than 255. Communityjev-tree walks a larger taxonomy one short Choice at a time and does not invent a leaf when the call fails. dabit3inbox-blitz packs seven,agent-assist nine, andturbo-rerank fifty Scores,jev-instant-search thirty-five questions, andjev-swarm up to eight agents, andjudge-sheets every column about one cell of source text, andJevTest measured 6.37× vs sequential on an 8-question probe — not official 193.6×.jev-demo-screener packs eight,jev-demo-rag twelve (six chunks × two), andjev-demo-logictest up to twelve Nouls in one call.jev-first-look saw 1 Noul median 143 ms vs 10 Nouls 157 ms after warm-up — author Austin client timings, not official Envoy process time.nak1b/jev-experiments puts 43 questions on one flight reading.bo7/jev_test asks two independent Choices (priority + category) — noteddydong/JevTest.QA-Classifier-Jev asks one Fact / Definition / Reason ChoicePRESS. Official Score page: at least two levels; the API accepts up to 10. CommunityTangerg/typesafe-sdk-go checked the live API on 2026-09-20 and saw a one-level Score accepted (always zero).fgn/jevgo documents the same one-level accept via OpenAPIminItems: 1 (official Python accepts; official TypeScript rejects).withzombies/jev-go follows the live minimum too. Prefer official docs on conflict.PRESSHow to write them is the next lesson.

What you do not get back

Jev does not return essays, code, tool calls, or chain-of-thought prose. Official docs: “jev-1.13 is not trained to generate text.” If you need an explanation for a human, call an LLM after (or beside) the decision—or show your own UI copy. Communityagent-assist never writes the reply (it picks a macro id) andjev-voice-turn never generates speech — audio stays on device.JevChat is a community experiment that still builds each word as a Choice — it is not official generation.uspraveen/Jevify,fidecastro/jevify, andchakuho expose a /v1/systemone surface over open checkpoints — they are not TypeSafe weights or RLCD.qwen-rlcd andLFM2.5-350M-RLCD put RLCD in the name only — one is an unbuilt-head prefix-fork, the other keeps stock LiquidAI weights.the 2.6B card,harshatheg, andGemma E2B do the same: pretrained scoring, not TypeSafe RLCD.eve-rlcd is the author’s own loop and says TypeSafe published no method.typesafe-jev-dojo falls back to amber SIM (local decisions) when there is no key — that SIM is not TypeSafe either.jev-lab without a key is a keyword heuristic that only mimics the response shape — also not Jev.jev-spec README “0.85 means empirically true in 85% of cases” is not TypeSafe docs — official: a high-confidence answer can still be wrongPRESS.

Why this feels more like code

Your application defines the answer space before inference. That constraint is the product: TypeSafe claims the model never makes type errors; schema matching is guaranteed by constructionCompany claim. You still design for wrong decisions via confidence and escalation. Communityrh-guard is the same split on a hook: code owns a structural deny; Jev is skipped after that fire; a soft score cannot independently create or cancel it.jevgate is the complementary envelope: the allowlist can prove a read; Jev only sees unknown verbs; the tool cannot blockPRESS.

Constraints worth knowing

• Context cited in models.md: 64k for state + all questions; 32k for state + longest question. Communityjev-analysis saw 400 max_tokens_exceeded from 36k on one path, and official ~100 ms as Envoy process time — not Oregon e2e.jev-demo-screener caps combined JD+resume at 50,000 characters for that 64k window.semantic-microscope saw Envoy p50 98 ms vs a Bengaluru client p50 356 ms — most of the trip was RTT, same lesson as jev-analysisPRESS
• English is primary; CJK and other languages are accepted with lower accuracy
• No customer fine-tune / LoRA; same weights for all accounts
• Not trained on customer requests or responses (official models.md)
• State is data, not a hostile user by default — add your own injection defenses

POST https://api.typesafe.ai/v1/systemone
Authorization: Bearer $TYPESAFE_API_KEY

# Python: pip install typesafe-sdk
# JS:     @typesafe-ai/sdk
# Pin:    jev-1.13.0  (aliases jev-latest / jev-preview can move)