Source: TypeSafe’s Jev 1.14 release post (ai-research/typesafe-news-jev-1-14-release-2026-10-14.md, 2026-10-14) for release dates; TypeSafe’s official documentation at docs.typesafe.ai, 28 pages fetched 2026-11-01 and saved one per file as ai-research/typesafe-docs-*.md (introduction, quickstart, coding agents, AI primer, System One, State, How to build, Use-case map, Primitives + Choice/Score/Noul/Box/Advanced, Images, Confidence, Patterns + four pattern pages, Models, Jev 1.13 jaggedness, Agent skill, API System One / List models / Errors, Cookbooks index).

Project: TypeSafe (Jev) · URL: https://docs.typesafe.ai/introduction · Version: jev-1.14.0 current (released 2026-10-14); jev-1.13.0 still served

The first-party reference for Jev, TypeSafe’s “System One” decision model. You send a state (text, JSON, or in beta, images) and any number of typed questions. Jev returns typed answers with probability distributions, and some types also return a confidence value. It never generates text. This page collects what the docs specify: the four question types, the confidence formulas, models and limits, the image beta, the integration patterns, and TypeSafe’s own list of where Jev 1.13 fails. The use-case and creator-test coverage stays in the main Jev article.

Key Takeaways

  • Four question types: Choice (one of up to 255 options), Score (a position on 2–10 described levels), Noul (probability that a yes/no statement is true), and Box (where something is in an image; beta, Jev 1.14 only). Answers are always confined to the options you supplied.
  • Unlimited questions per request. The only bound is context: 64k tokens per request and 32k per question. Questions run in parallel and independently, so “asking a question you might not need is close to free”. Most calls take “about 100 ms”.
  • Price unchanged: 42 per billion), output free. The same price applies to 1.13 and 1.14, to long context, and to image tokens.
  • Confidence is a published formula, not a black box. Choice: (p_max − 1/n)/(1 − 1/n). Score: a distance-weighted spread measure. Noul has no confidence: the probability is the uncertainty. Thresholds are your code’s job, set per action by risk.
  • Images are a “Dreamer-ready” beta, off by default since the 1.14 release (2026-10-14). Turn them on with the TypeSafe-Danger-Zone-Images: true header. The System One and Models pages still say “text only”. They predate the release post (ai-research/typesafe-news-jev-1-14-release-2026-10-14.md), which announced images, Box and long context as experimental.
  • Long context is opt-in: the TypeSafe-Danger-Zone-Long-Context: true header raises both limits to 262k tokens at the same price, but “slower and less accurate”.
  • TypeSafe lists nine failure modes for 1.13: literal reading, maths and counting, dates, indirection, large irrelevant state, adversarial content (injected text “can move the answer”), contradictory criteria, a bias toward the first Choice option, and text generation.
  • It is not a coding-agent model. “There is no model: "jev-latest" setting that turns your coding agent into a Jev-powered agent.” Use your coding agent to write code that calls Jev.
  • Not customisable by fine-tuning. The same weights serve every account. You adapt it through state, instructions and criteria. TypeSafe says it doesn’t train on customer requests, and offers zero data retention for enterprise.

How It Works

  • System One: the name is from Kahneman’s Thinking, Fast and Slow: fast, intuitive judgment rather than deliberate reasoning. TypeSafe’s design rule is that each question should be “a gut-check determination… a highly knowledgeable person could make in a few seconds.”
  • RLCD training: “reinforcement learning for calibrated decisions”, positioned as a third post-training path after RLHF (chatbots) and RLVR (reasoning models). Calibrated means outcomes given 0.8 should happen about 80% of the time, measured across groups of predictions, never guaranteed for one answer. The docs credit cofounder Diogo Almeida with co-inventing RLHF.
  • TypeSafe’s thesis (“Machine Native Intelligence”, “Building prod, not God”): large-scale automation will be “closer to 99% machine-to-machine interactions and 1% human interaction”. So the output contract should suit software, not readers. RLHF rewards sycophancy and causes “mode dropping”.
  • Architecture TypeSafe recommends: “AI-powered software”, where code owns control flow and the model handles narrow semantic judgments. It is contrasted with agents, where “every loop introduces another opportunity to go off the rails.”

The Four Question Types

TypeUse whencriteriaAnswer fieldsNotes
ChoiceOne of a fixed, unordered setMap of option → description (up to 255)choice, probabilities (sum 1), confidenceAdd other/none, or better, a companion Noul “does anything fit?”
ScoreA position on a spectrumOrdered array of 2–10 level descriptions (low → high)score (probability-weighted mean, can be fractional), legend, probabilities, confidenceEach level is judged on its own; the model never sees level numbers
NoulYes/no, where the probability is the useful signalOptional {true, false} descriptionsnoul (0–1)No confidence. Start at 0.7 yes / 0.3 no; send the middle band to a person
Box (beta)Where something is in an imageNonebox {top,bottom,left,right} as fractions; confidenceAlways returns a box (no “not found”): pair it with a Noul. One target per question. Needs jev-1.14

Writing questions:

  • Question IDs aren’t sent to the model. Put the full question in instructions.
  • Point at fields with backticked paths, e.g. “Does ticket.messages[0].text request a refund?”. instructions and criteria accept strings, objects or arrays (up to 64 nesting levels).
  • Score levels must describe situations, not degrees. In TypeSafe’s example, numbers-only levels ("0","1","2") scored a cosmetic bug at 0.55 with confidence 0.33. Descriptive levels scored it 0.0 at confidence 1.0.
  • Examples help only if they resemble real inputs. A relevant example moved one answer from score 1.43 / confidence 0.35 to 1.03 / 0.96, and an unrelated example changed nothing. Higher confidence “does not establish which answer is correct”. Validate on held-out inputs.
  • Noul is not a scale. “Is the candidate strong in Python?” at 0.81 means “81% likely yes”, not “81% skilled”. Use a Score for degree.
  • One condition per question. “Angry and asking for a refund” should be two Nouls combined in code. Phrase questions so a high value means yes.

Confidence

  • Choice: confidence = (p_max − 1/n) / (1 − 1/n), where 0 is an even split and 1 is certain. Only the top probability counts, so (0.6, 0.3, 0.1) and (0.6, 0.2, 0.2) both give 0.4.
  • Score: 1 − (Σ p_i·|i − m|) / MAD_unif, floored at 0. Splitting probability between neighbouring levels costs less confidence than splitting between the ends: with three levels, (0, 0.5, 0.5) gives 0.25 and (0.5, 0, 0.5) gives 0.
  • Noul (optional): use |2p − 1| if you want one gate for both Nouls and Choices.
  • Alternatives the docs suggest: the raw top probability (set its threshold per question), or the top-to-second ratio for decisions that come down to two candidates.
  • Usage: three bands: act, confirm, escalate. Thresholds rise with how costly a wrong action is. TypeSafe’s voice-banking example uses a 0.6 floor for everything, then requires more than 0.85 before approving a transfer without confirmation. Plot confidence against accuracy on your own data before you trust a threshold.

Models, Limits and Access

Jev 1.14Jev 1.13
Model IDjev-1.14.0 (from 2026-10-14)jev-1.13.0
Price$0.042 / Mtok input, output freeSame
Context64k per request / 32k per questionSame
Long context (header)262k / 262k, same price, slower and less accurateSame
Images (beta header)Yes, plus BoxYes, no Box
  • Aliases: jev-latest (the stable release and SDK default) and jev-preview. Both point to jev-1.14.0 today. Aliases move on release, so pin the version ID once you’ve tuned thresholds. Every response’s model field shows which version answered.
  • Endpoint: POST https://api.typesafe.ai/v1/systemone with a Bearer key from console.typesafe.ai. GET /v1/models lists the aliases.
  • SDKs: Python typesafe-sdk (Python ≥ 3.10), JavaScript @typesafe-ai/sdk (Node ≥ 20; it infers answer types from your questions). Both retry 429/529 with backoff by default.
  • Request rules: 32 MiB body, gzip or zstd allowed, finite numbers only (send IDs as strings). Status codes: 400, 401, 413, 415, 422 (validation), 429 (rate limit), 529 (overloaded).
  • Rate limits are per organisation, in tokens/s and requests/s, shown in the console. Raise them through a request form or a custom/enterprise plan.
  • Languages: English is primary. Other languages, including CJK, work at lower accuracy.
  • Playground: console.typesafe.ai/playground, for testing with no code.

Images (Beta)

  • Turning it on: send TypeSafe-Danger-Zone-Images: true on every request that has an image or a Box question. Without it you get a 400.
  • Sending images: list each one once in a top-level assets map under a reference string (use a fresh UUID; references are matched as plain text anywhere, so cat would also match inside category). Write the reference wherever the image belongs: in the state, a question’s instructions, or criteria. Choice options can be defined by example pictures.
  • Upload: multipart is recommended over base64, which is about a third larger. The SDKs create references for you.
  • Limits: PNG or still JPEG only. Each side 280–1568 px. 16 MiB per file and 32 MiB per request, counting every reference expansion. Every use of a reference adds the image again, so put shared images in the state once.
  • Prep: the API strips EXIF and does not apply orientation, so rotate the pixels yourself. Resize to a longest side of 1568 px, re-save progressive JPEGs as baseline, and convert GIF, WebP or HEIC.
  • Billing: images bill as input tokens at the same rate. TypeSafe has promised guidance on image token estimates but hasn’t published it.
  • Patterns: compare against a known picture, define options by example, “what changed” across frames, look closer only when needed, tile search on large images, Box for location.

Build Patterns

PatternIdeaTypeSafe’s example
Speculative fan-outAsk every question code might need in one call, ignore the irrelevant answersTicket triage: category + bug severity + repro steps + refund + frustration, one call
Confidence-gated routingThe answer says what, confidence says whether to actVoice banking: a 0.6 floor, a higher bar for transfers
Composite scoringSplit a judgment into atomic Scores, normalise each by len(levels) − 1, weight in codeResume ranking: 0.4·Python + 0.1·leadership + 0.4·design + 0.1·generalist
Intent routingJev classifies cheaply in front of the expensive handlersCustomer service: deterministic code, two specialist LLMs, or a human

TypeSafe’s eight build steps:

  1. Use code when you can.
  2. Send only relevant state.
  3. Structure the state.
  4. Decompose questions. TypeSafe calls this “probably the most important concept.”
  5. Structure the questions.
  6. Ask many questions per call.
  7. Combine answers in code, or feed the probabilities to a classical ML model as features.
  8. Route on uncertainty.

A second request is justified only when your code can’t build it until it has the first answer.

Where Jev 1.13 Fails (TypeSafe’s Own List, Reviewed 2026-10-02)

Failure modeTypeSafe’s fix
Literal reading: answers what you wrote, not what you meantState the exact condition; put boundary cases in criteria
Maths, counting, numeric representations (hex colours, RGB)Do it in code; send named buckets
Date and time comparisonExtract date parts as Choices (with a “not stated” option); compare in code
Indirection, double negatives, multi-hopAsk directly; name the relevant state fields
Large state full of irrelevant detail (“context rot”)Filter first, or pre-filter with Nouls
Adversarial content: injected instructions can move answersPrecise criteria; test edge cases before deploying
Instructions and criteria that disagreeAlign them in plain language
Choice option order: leans toward the first optionReorder options and check the answer holds
GenerationUse a generative model; turn extraction into a Choice over candidates

No equivalent page for 1.14 has been published.

Use Cases and Cookbooks

  • TypeSafe’s five categories: AI automation software, real-time apps (“150ms”), AI map-reduce over big data, universal verification (checking other AIs’ outputs, tool calls, citations, jailbreaks), and harness engineering (model routing, context retrieval, guardrails).
  • Industry lists include lead generation (match company profiles and inbound messages to an ICP, score fit and maturity), recruiting, scientific review, semantic code linting, and feature extraction for predictive models.
  • 25 cookbooks plus a demo, with measured results (all of them: Jev Cookbooks). Notable ones:
    • Re-ranking: BM25 shortlists on 40 CLERC legal queries; top-1 accuracy 5% → 18%, top-10 38% → 62%.
    • Parallel questions: 13 GDPR questions in one call vs 13 calls. The cookbook page itself reports 12.2× cheaper and 10.0× faster. The primitives page quotes 11.5× and 9.6×, probably from an earlier run.
    • Skill suggestion: picks one skill from Nous Research’s Hermes catalog (194 skills per the cookbook itself; the primitives page says 182) in two requests.
    • Classification using confidence: sorts SEC filings into 75 industry groups, falling back to the parent division when confidence is low.
    • Guardrails, citation checking, RAG-passage filtering, date extraction, OCR verification with Box, and Twitch chat filtering.

Agent Skill

  • Install in Claude Code: claude plugin marketplace add typesafe-ai/skills, then claude plugin install typesafe@typesafe-ai. Invoke it with /typesafe:typesafe-ai.
  • Install in other agents: npx skills add typesafe-ai/skills --skill typesafe-ai (project-local; add -g for global).
  • Updates: claude plugin marketplace update typesafe-ai and claude plugin update typesafe@typesafe-ai. A stale skill makes agents “invent request or response fields”.
  • TypeSafe’s review advice: keep all questions and thresholds as constants in one file. “Agents aren’t great at writing questions, so expect to edit collaboratively with them.” Coding agents also slip into one question per request, which the skill corrects.

Try It

  1. Paste a real inbound lead or patient message into the Playground. Ask one Choice (service line), one Score (urgency, with situation-style levels) and two Nouls (asks for pricing? existing patient?) in a single call, and compare the probabilities with your own call.
  2. Before trusting a Choice in production, run each test input twice with the option order reversed. The first-option bias is TypeSafe’s own finding.
  3. Pin jev-1.14.0 in code once thresholds are tuned. Re-tune before moving to whatever jev-latest becomes.
  4. Treat image questions as a prototype. Set the beta header, prepare images to 1568 px with upright pixels, and pair every Box with a “is it there?” Noul.
  5. Install the agent skill in Claude Code and run TypeSafe’s brainstorming prompt: “find opportunities for using intelligent judgement to stand in for complex parsing or other fragile code.”

Open Questions

  • The System One and Models pages still say “text only”. The 1.14 release post shows the Images and Box pages are the current ones, but TypeSafe hasn’t corrected the older pages.
  • No jaggedness page exists for 1.14, and no changelog says what changed from 1.13 to 1.14 besides Box.
  • Image token costs aren’t published (“guidance… soon”).
  • Latency figures vary by page (“about 100 ms” in How to build, “150ms” in the use-case map) and aren’t benchmarked in the docs.