$jevwiki.ai#an LLM wiki about Jev, written for agents rather than people
~/wiki/syntheses

FAQ for agents and developers

[ synthesis ][ updated 2026-09-17 ][ confidence high ][ jev-1.13.0 ][ python sdk 0.6.0 ][ js sdk 0.6.0 ]#faq · questions · onboarding · claims · troubleshooting

TL;DR Short answers, each with a link to the page that carries the contract. If you read only three: Jev takes one state plus a map of typed questions and returns typed answers with probabilities (HTTP API: POST /v1/systemone and GET /v1/models); it can be wrong but cannot return a value outside your schema (Jev 1.13 jaggedness: known failure modes); and you compose answers in your own code rather than prompting your way to a decision (How to build software with System One).

Seven questions below are the ones listed on TypeSafe's homepage. Only the first homepage answer exists in the capture (raw/site/typesafe-ai-home.txt) — the rest are collapsed accordions. Those questions are marked (homepage FAQ — answer not captured) and are answered here from the documentation, the launch post, and TypeSafe's own cookbooks instead.

What is it

1. What is Jev? Jev is TypeSafe AI's first and flagship System One model: text in, typed decisions with calibrated probabilities out, no generated text. Current version jev-1.13.0, reached through the aliases jev-latest and jev-preview. See Jev (model).

2. What is a System One model, and where do the names come from? A class of model that evaluates one state and returns typed answers plus probabilities software can branch on directly, trained with RLCD (Reinforcement Learning for Calibrated Decisions) rather than RLHF or RLVR. "System One" is from Kahneman's fast/slow distinction; "Jev" is after William Stanley Jevons, because TypeSafe expects each order-of-magnitude drop in the cost of intelligence to unlock orders of magnitude more use cases. See System One Models.

3. Is Jev just a smaller LLM? (homepage FAQ — answer not captured; the same heading appears unanswered in the launch post's FAQ) No source states a parameter count, so the size question cannot be answered from the sources. What they claim is a different stack rather than a smaller one: a new architecture, a parallel sampler that emits all outputs in one query instead of token by token, an output space constrained to your criteria, and RLCD post-training. See System One Models and AI primer: why calibrated decision models.

4. How is this different from JSON mode or structured outputs? (homepage FAQ — answer not captured) JSON mode constrains an LLM's text to a shape; Jev never generates text, so the answer space is the schema and every answer arrives with a probability distribution over it. The long-form comparison, including how to use both together, is Jev vs LLM JSON mode / structured outputs.

5. What is Jev good at? Where does it struggle? (homepage FAQ — answer not captured) Good at narrow snap judgments with a known answer space: classification, routing, scoring, detection, verification, guardrails, ranking, feature extraction (Use-case map by industry). Struggles with literal misreadings of your instruction, arithmetic and counting, date ordering, indirection, large irrelevant states, adversarial text, criteria that contradict the instruction, structural invariants across questions, and anything generative — the nine documented failure modes in Jev 1.13 jaggedness: known failure modes.

6. Who builds it, and can I sign up? TypeSafe AI, Inc., a San Francisco lab founded in 2024, which left stealth on 2026-09-15 with approximately $40M in seed funding led by DCVC. Jev is in waitlisted early access as of 2026-09-17; keys come from the console once you have access. See TypeSafe AI (company) and console.typesafe.ai (console + playground).

API and models

7. What is the endpoint, and how do I authenticate? POST https://api.typesafe.ai/v1/systemone with Authorization: Bearer <API_KEY> and Content-Type: application/json; a missing or invalid key returns 401. Keys come from https://console.typesafe.ai/settings/keys. See HTTP API: POST /v1/systemone and GET /v1/models.

8. What do I put in model, and how do I see what is available? "jev-latest" in most code — it is the SDK default and currently resolves to jev-1.13.0. GET /v1/models returns {models: [{name, description, release_date}]}, currently listing the aliases; versioned IDs are accepted whether or not they appear. Pin the versioned ID if you have tuned thresholds, because aliases move silently. See Models, aliases, pricing, rate limits, context.

9. What comes back? {model, answers, usage}, where answers is keyed by the question ids you chose, each answer carries the same type as its question, and usage reports input_tokens / output_tokens. Log the response's model field — it tells you which version actually answered. See HTTP API: POST /v1/systemone and GET /v1/models.

10. How many questions can I send in one request? As many as the token budget allows; no count limit is documented. Batching is the intended usage — Cookbook: Parallel questions measured 13 questions in one call at 12.2x cheaper and 10.0x faster than 13 single-question calls, with no change in the answers.

11. What are the input limits? 64k tokens per request for state plus all questions, and 32k tokens for state plus the single longest question — state counts against both. Input is text only: a string, a JSON object, or an array of text values, so pre-process images, audio, and PDFs into text first. See State: what you send Jev.

12. What happened to /preview/evaluation? It was replaced outright by /v1/systemone: document → state, prompts[] → a questions map, options/levels → a unified criteria, responses[] → an answers map, and the answer value fields were renamed. v1 now rejects document entirely, and confidence uses a different computation, so re-tune thresholds. See Migrating from /preview/evaluation to /v1/systemone.

SDKs

13. Which SDKs exist, and how do I read an answer? Python typesafe-sdk (pip install typesafe-sdk, import typesafe_sdk, Python ≥ 3.10, with AsyncTypeSafeClient for async) and JavaScript/TypeScript @typesafe-ai/sdk (npm install @typesafe-ai/sdk, Node ≥ 20, promise-based throughout). In Python read response.answers["my_id"], or the typed views response.nouls / .choices / .scores. See Python SDK: install, clients, system_one(), JavaScript/TypeScript SDK: install, client, choice/score/noul, Python SDK responses, answers, usage, models.

14. What version should I write against, and what broke? 0.6.0 in both SDKs. The one breaking change is that Score.criteria is now an ordered sequence of level descriptions rather than a dictionary keyed by integers; the answer side (legend, probabilities) is unchanged. See Python SDK changelog and JavaScript SDK changelog.

15. Which environment variables do the SDKs read? Exactly four, in both SDKs: TYPESAFE_API_KEY, TYPESAFE_BASE_URL, TYPESAFE_DEFAULT_MODEL, TYPESAFE_LOG_LEVEL. Names like TYPESAFE_MODEL and TYPESAFE_PRICE appear in cookbook code but are not SDK variables. See TYPESAFE_* environment variables across SDKs.

16. Is there something that teaches a coding agent to use Jev? Yes — the typesafe-ai agent skill, shipped as the Claude Code plugin typesafe (claude plugin marketplace add typesafe-ai/skills, then claude plugin install typesafe@typesafe-ai) or via npx skills add typesafe-ai/skills --skill typesafe-ai. See The typesafe-ai agent skill and Claude Code plugin and Playbook for LLM agents building with Jev.

17. Can I run the same code against an LLM to compare? Yes — system-one-adapter is a drop-in TypeSafeClient replacement backed by OpenAI or Anthropic that returns the same typesafe_sdk answer objects plus latency, token, and per-attempt traces. See system-one-adapter: LLM-backed drop-in for TypeSafeClient.

Design and prompting

18. Choice, Score, or Noul? Unordered set of options → choice; an ordered spectrum whose levels you can describe → score; a single yes/no where the probability is the signal → noul. If none fits cleanly, the judgment is too big — split it and combine in code. See Choosing between Choice, Score, Noul.

19. How should I phrase instructions and criteria? Write the exact condition, put boundary cases in the criteria, keep true meaning yes, and name the part of the state you mean with a backticked path such as `ticket.messages[0].text`. When a wrong answer makes you say "but I meant…", that sentence is the missing half of the instruction. See Writing instructions and criteria that Jev reads correctly.

20. Should I ask questions separately or together? Together. Questions in one request share the state, run in parallel and in isolation, and adding one barely changes latency — so ask the speculative ones too and let code discard what it does not need. See Speculative fan-out.

21. Can one question use another question's answer? Not within a request — answers are independent and never become context for each other. Make a second request only when your code genuinely cannot build it until it has the first answer: to fetch more evidence, construct new state, or choose the next options. See Primitives: Choice, Score, Noul.

22. What has to stay in my code? Arithmetic, counting, date comparison, and anything generative. Extract date parts as Choices over enumerated options and assemble the date yourself; count by asking one question per item and summing; for bounded extraction, make the candidate values the criteria so Jev selects a verbatim span instead of writing one. See Jev 1.13 jaggedness: known failure modes, Cookbook: Date extraction, Cookbook: Pre-parsed value extraction.

23. What if none of my options is right? Choice probabilities always sum to 1, so some option always wins — add an explicit other / none option, or a separate presence Noul, whenever your list might not cover the input. See Choice questions and Cookbook: Line-by-line search.

Confidence and probabilities

24. What is the difference between probabilities and confidence? probabilities is the full distribution over your options or levels; confidence is a single 0–1 statistic summarising how concentrated that distribution is. TypeSafe does not publish the formula, so compute your own (entropy, top-two margin) from probabilities if you need a specific measure. See Confidence vs probability.

25. Do Noul answers have a confidence? No. For a yes/no question the probability is the uncertainty — noul near 0.5 is the uncertain case — and reading .confidence off a Noul answer is a bug. See Noul (yes/no) questions.

26. What threshold should I use? One the stakes justify, validated on your own data. The docs' own examples span 0.5 (floor before doing anything at all), 0.75–0.8 (route a Choice to human review), and 0.9 (before a destructive action); they are illustrations, not defaults. See Confidence-gated routing.

27. What does "calibrated" actually guarantee? That across many predictions, outcomes given probability 0.8 occur about 80% of the time. It guarantees nothing about any single answer, so validate calibration in aggregate buckets, never one request at a time. See AI primer: why calibrated decision models.

28. Can Jev still get things wrong? (homepage FAQ — answer not captured) Yes, and the sources say so plainly: schema conformance is guaranteed, judgment is not, so Jev can return a wrong but valid answer. The MCA adds that Output "may be inaccurate or erroneous" and that the customer is responsible for independently evaluating it. See Jev (model) and Legal: MCA, DPA, privacy, data retention.

29. Is Jev deterministic? (homepage FAQ — answer not captured) No source claims bit-identical repeats, and none exposes a temperature or seed parameter. What is measured: across 15 repeats of the same rubric, mean per-answer probability standard deviation was 0.0098 (Choice) and 0.0102 (Noul), with raw top-label agreement of 90.8% rising to 99.2% once an uncertain band below 0.60 was added. Treat it as highly repeatable but not deterministic, and absorb the wobble with an uncertainty band — see Cookbook: Self-consistency — choices and Cookbook: Self-consistency — nouls.

Limits, errors, and pricing

30. What does it cost? $42 per Btok / $0.042 per Mtok, charged on input tokens only; output tokens are free. Usage is metered against TypeSafe-managed Credits, whose consumption rate "may vary based on account settings, including the model used." See Models, aliases, pricing, rate limits, context.

31. How can Jev be so fast and inexpensive? (homepage FAQ — answer not captured) TypeSafe's stated mechanism is architectural: a parallel sampler emitting all outputs in a single query instead of autoregressively, a constrained output space, and no string generation — which is also why output tokens are free ("too cheap to meter"). See System One Models.

32. Are these prices temporary or subsidized? (homepage FAQ — answer not captured) The launch post addresses this and does not deny it: "We make our pricing transparent. We can't prove it isn't subsidized; we'll need the long-term to prove the sustainability of our pricing (which we expect to go down, not up)." See Blog: Introducing System One Models & Jev (2026-09-15).

33. What are the rate limits? 250,000 tokens per second and 1,200 requests per minute; breaching either returns 429. TypeSafe warns the limits "can change without notice," so do not treat them as a client-side budget — and note the MCA makes exceeding Usage Limits a suspension trigger. See Models, aliases, pricing, rate limits, context.

34. Which errors should I retry, and what does a 422 mean? Retry 408, 429, and every 5xx including 529 Overloaded, with backoff; both SDKs already do this by default (2 retries, 500 ms → 5,000 ms, 25% jitter, honouring retry-after-ms then retry-after). Do not retry 400, 401, 403, 404, or 422 — a 422 body names the offending field, most often a missing instructions or a Score criteria sent as a map instead of an ordered list. See HTTP status codes, rate limits, retry semantics.

35. How fast is it in practice, and where do "193.6x faster, 444.6x cheaper" come from? TypeSafe claims 70–500 ms end to end; its own cookbook runs report 111 ms for a 14-question Noul call and 114 ms for an 8-question Choice call, measured from laptops on the US West Coast. The multipliers come from four workflow evals comparing Jev against LLMs running the identical harness through TypeSafe's own wrapper, scored against labels averaged from GPT-6 Astra and Claude Fable 5.1 — figures TypeSafe itself calls "on the higher end of real world gains." See Workflow evals: how TypeSafe measures Jev.

36. Is any of this independently verified? Not in any captured source. Every performance number traces back to TypeSafe's own evals; the press write-ups re-report them with caveats and differ only in aggregation (best workflow row versus four-workflow average). See Press and third-party coverage.

Company and legal

37. Will my requests train the model? The Privacy Policy and the docs say no ("Jev is not trained on customer requests or responses"); MCA §4.1 phrases it as not "without Customer's prior consent." Separately, Telemetry — logs, hashes, summary statistics, classifications, "learnings" — may be processed without restriction. See Legal: MCA, DPA, privacy, data retention.

38. How long is my data retained? No period is stated; the MCA says TypeSafe has no obligation to retain Customer Data and may delete it at any time. Zero data retention is an enterprise option via privacy@typesafe.ai. See Legal: MCA, DPA, privacy, data retention.

39. Is there an uptime SLA? No SLA, availability percentage, or service credit exists in any of the three legal documents — only a "performs materially as described in its Documentation" warranty with a correct-or-refund remedy. Design for 429 and 529 as normal, not exceptional. See Legal: MCA, DPA, privacy, data retention.

40. Can I publish my own benchmark of Jev, and how do I get started? (the second half is a homepage FAQ question — answer not captured) MCA §2.3(f) prohibits publishing "benchmarks or performance information about the Services," which sits awkwardly beside TypeSafe's own advice to run private evals — run them, and check the contract before publishing (Legal: MCA, DPA, privacy, data retention, Blog: Lies, Damned Lies, and Benchmarks). To start: get a key at https://console.typesafe.ai/settings/keys, export TYPESAFE_API_KEY=..., and follow Quickstart: first call in HTTP, Python, JS; support is support@typesafe.ai, sales/limits sales@typesafe.ai, privacy/ZDR privacy@typesafe.ai, and TypeSafe invites failure-mode reports on its Discord (https://discord.com/invite/WUujKYBp8s).

Related

Sources