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

Jev Wiki — Master Index

TL;DR 88 content pages (plus this index and the log) about Jev, TypeSafe AI's System One model. Snapshot of upstream sources: 2026-09-17. Model: jev-1.13.0; typesafe-sdk 0.6.0; @typesafe-ai/sdk 0.6.0. New here? Read Playbook for LLM agents building with Jev first, then HTTP API: POST /v1/systemone and GET /v1/models or the SDK reference, then Jev 1.13 jaggedness: known failure modes.

This file is generated from page frontmatter by node scripts/build.mjs; edit the pages, not this list. Maintenance log: Activity log.

Guides (6)

Page Summary ~Tokens Updated Conf.
Playbook for LLM agents building with Jev Seven-step playbook for an agent told 'use Jev for X': decide, decompose, shape state, write questions, call, consume, handle errors, test. 7.0k 2026-09-20 high
Choosing between Choice, Score, Noul Decision table mapping question shape to Choice, Score or Noul, the anti-patterns for each, and how to decompose a complex judgment into several primitives. 3.4k 2026-09-17 high
Quickstart: first call in HTTP, Python, JS Get a key at console.typesafe.ai, export TYPESAFE_API_KEY, and make your first POST /v1/systemone call in curl, Python, or TypeScript. 3.6k 2026-09-17 high
Smart home assistant demo walkthrough TypeSafe's smart home demo: one speculative fan-out per utterance routes device commands, while an LLM handles compound splitting and conversational fallback. 2.3k 2026-09-17 medium
Testing and evaluating a Jev workflow Measure accuracy, calibration, repeatability, latency and cost on your own labelled set; pick thresholds from confidence buckets; regression-test jaggedness. 4.3k 2026-09-21 high
Writing instructions and criteria that Jev reads correctly How to phrase instructions and criteria for each question type, with before/after rewrites for literal reading, negation, numbers and dates, plus an edge-case testing loop. 4.6k 2026-09-17 high

Ideas (community) (12)

Page Summary ~Tokens Updated Conf.
Community repos: what people built and how they use Jev Index of 58 third-party Jev repos, one line each, split across two domain pages, plus reusable designs and a vetting checklist. 2.6k 2026-09-21 medium
Consult guide: could Jev help this project? Procedure for an agent asked 'here is my project, could Jev help?': what to read (10-15k tokens), how to find and filter decisions, and the shortlist format to return. 2.0k 2026-09-21 medium
Field reports: independent evaluations, critiques, open replicas What independent testers measured in Jev's first week: repeatable bounded decisions, weak standalone browser control, and probe-measured order and option-set sensitivity. 3.4k 2026-09-21 medium
Measurements, access routes and open replicas Numbers measured with Jev — cost, latency, calibration, token accounting, vendor harness figures — plus access routes and replica trust notes. Every figure is one run; re-measure yours. 3.0k 2026-09-21 medium
Ideas section overview One-screen map of the community-tier ideas section: which index or domain page to open for which question, and how far to trust them. 0.8k 2026-09-21 medium
Decision patterns from the community (with fit verdicts) Index of all 37 community decision patterns (ID, primitive, fit verdict, page), routing by project type to one of four domain pages, retrofit order, and anti-patterns. 2.4k 2026-09-21 medium
Patterns: agent internals, context and coding agents P01-P11 plus P35-P37: community decision patterns for agent loops, model and security routing, context, memory, coding agents and self-compiling workflows, each with a fit verdict. 4.4k 2026-09-21 medium
Patterns: marketing, sales, GTM, content, support and ops Eleven community decision patterns for GTM, content, support and ops — lead scoring, next-best action, churn, draft QA, SEO, triage, approvals — each with a fit verdict. 4.5k 2026-09-21 medium
Patterns: judging, search, documents, real-time and markets P15-P18 and P24-P27: community decision patterns for judging, moderation, search and ranking, corpus labelling, documents, transcripts, devices and markets, with fit verdicts. 3.0k 2026-09-21 medium
Patterns: browser, computer use, voice and product UI Community patterns where Jev picks the next on-screen action or labels items inside a product UI: browser agents, site-exposed tools, desktop/mobile/voice control, in-UI classifiers (P12-P14, P34). 1.3k 2026-09-21 medium
Repos: coding agents, orchestration, memory, browser and computer use, integrations 32 community repos that put Jev inside a coding agent, an orchestrator, a browser or computer-use loop, or an MCP/CLI integration surface. 4.4k 2026-09-21 medium
Repos: data, documents, judging, real-time, markets, business apps, replicas 23 community repos that point Jev at rows, documents, media, live simulations, markets, business questions — plus the open replicas of the interface. 3.8k 2026-09-21 medium

Concepts (13)

Page Summary ~Tokens Updated Conf.
Structured instructions, options, levels, criteria instructions, Choice option descriptions, Score levels and Noul true/false all accept JSON: string, object, array or null (EntryType). 3.4k 2026-09-17 high
Choice questions Choice picks one option from a fixed set: send type/instructions/criteria, get back choice, probabilities over every option, and confidence. 4.6k 2026-09-17 high
Confidence vs probability confidence is a 0-1 statistic derived from an answer's probabilities distribution; Choice and Score carry it, Noul does not. 2.3k 2026-09-17 high
How to build software with System One Build a normal software workflow and insert System One only where AI is needed: narrow typed questions, composed in code. 6.3k 2026-09-17 high
Jev 1.13 jaggedness: known failure modes Nine documented failure modes of jev-1.13 — literal reading, math, dates, indirection, big states, adversarial text, contradictions, invariants, generation — with the mitigation for each. 3.3k 2026-09-17 high
AI primer: why calibrated decision models Why TypeSafe post-trains with RLCD for calibrated decisions instead of RLHF/RLVR for generated text. 1.7k 2026-09-17 high
Noul (yes/no) questions Noul asks one yes/no question and returns a single number, the probability that the answer is yes; criteria are optional true/false clarifications. 2.3k 2026-09-17 high
Primitives: Choice, Score, Noul The three Jev question types (Choice, Score, Noul), the typed answers they return, how to pick one, and how to batch many questions in one request. 3.0k 2026-09-17 high
Score questions Score rates a state against an ordered array of described levels and returns a probability-weighted position, the legend, per-level probabilities, and confidence. 5.1k 2026-09-17 high
State: what you send Jev State is the content Jev evaluates: a string, JSON object, or array of text, shared by every question in one request. 1.7k 2026-09-17 high
System One Models System One models make fast, typed, calibrated decisions software can consume directly; Jev is TypeSafe's first one. 2.8k 2026-09-17 high
Use-case map by industry Scannable map of where Jev fits: five capability categories, eighteen industry/function areas, and ten decision shapes. 3.2k 2026-09-20 high
Workflow evals: how TypeSafe measures Jev TypeSafe's four workflow evals: same harness for every model, reference labels from GPT-6 Astra + Claude Fable 5.1, source of 193.6x/444.6x. 4.0k 2026-09-17 high

Reference (18)

Page Summary ~Tokens Updated Conf.
The typesafe-ai agent skill and Claude Code plugin Install, update, and invoke the typesafe-ai agent skill (Claude Code plugin typesafe 0.5.7, MIT), plus a faithful digest of the guidance SKILL.md gives agents. 5.2k 2026-09-21 high
TYPESAFE_* environment variables across SDKs Exactly four TYPESAFE_* environment variables are read by the SDKs: API_KEY, BASE_URL, DEFAULT_MODEL, LOG_LEVEL — both SDKs read all four. 2.8k 2026-09-17 high
HTTP API: POST /v1/systemone and GET /v1/models Complete wire contract for POST https://api.typesafe.ai/v1/systemone and GET /v1/models: auth, request body, question and answer shapes, usage, errors. 5.0k 2026-09-17 high
JavaScript/TypeScript SDK: install, client, choice/score/noul @typesafe-ai/sdk 0.6.0: install, TypeSafeClient config and defaults, systemOne(), choice/score/noul builders, APIPromise, models.list(), logging. 6.9k 2026-09-17 high
JavaScript SDK changelog @typesafe-ai/sdk releases: 0.5.7 initial public release, 0.6.0 (2026-09-15) makes Score.criteria an ordered sequence — the one breaking change. 1.8k 2026-09-17 high
JavaScript SDK error classes, RetryPolicy, RequestOptions Error hierarchy of @typesafe-ai/sdk 0.6.0, status-to-class mapping, RetryPolicy defaults (2 retries, 500ms/5s backoff, 0.25 jitter) and RequestOptions. 5.2k 2026-09-17 high
JavaScript SDK interfaces and type aliases Every interface, type alias and exported variable in @typesafe-ai/sdk 0.6.0, with per-property type/required/default tables and generics examples. 6.0k 2026-09-17 high
Legal: MCA, DPA, privacy, data retention Developer-facing summary of TypeSafe's MCA, DPA, and Privacy Policy: no training on Input, telemetry carve-out, retention, sub-processors, acceptable use, no uptime SLA. 5.1k 2026-09-17 high
Migrating from /preview/evaluation to /v1/systemone Every delta between the preview evaluation endpoint and the stable v1 API, with before/after examples, plus the typesafe-client to typesafe-sdk rename table. 3.7k 2026-09-17 high
Models, aliases, pricing, rate limits, context Jev 1.13 model ID, aliases, $42/Btok pricing with free output tokens, 250k tok/s and 1,200 rpm limits, 64k/32k context rule, and text-only input. 2.7k 2026-09-17 high
OpenAPI component schemas Every component schema in TypeSafe's OpenAPI 3.1.0 document (info version 0.2.0): properties, types, requiredness, constraints, and examples. 4.8k 2026-09-17 high
Python SDK: install, clients, system_one() typesafe-sdk 0.6.0: install, TypeSafeClient/AsyncTypeSafeClient constructor params, system_one() kwargs, models resource, env vars, logging, and the full export list. 5.7k 2026-09-17 high
Python SDK changelog typesafe-sdk release history: 0.0.1a0 placeholder, 0.5.7 initial public release (2026-09-14), 0.6.0 (2026-09-15) with the Score.criteria breaking change. 1.8k 2026-09-17 high
Python SDK question types (Noul, Choice, Score) Every field of Noul, Choice and Score in typesafe-sdk 0.6.0, their TypedDict equivalents, JSONContent typing, client-side validation, and the 0.6.0 Score.criteria breaking change. 3.1k 2026-09-17 high
Python SDK responses, answers, usage, models SystemOneResponse fields and the .nouls/.choices/.scores views, every answer attribute, Usage, request_id and raw_http_response, plus ListModelsResponse and ModelMetadata. 3.4k 2026-09-17 high
Python SDK retries, exceptions, constants RetryPolicy fields and defaults, which statuses and exceptions are retried, backoff and Retry-After handling, the full exception hierarchy with attributes, timeouts, and public constants. 4.7k 2026-09-17 high
HTTP status codes, rate limits, retry semantics One table mapping every HTTP status to its meaning, Python exception, JS error class, retryability, and the recommended client action. 3.6k 2026-09-17 high
system-one-adapter: LLM-backed drop-in for TypeSafeClient system-one-adapter answers typesafe_sdk questions with an OpenAI or Anthropic LLM instead of Jev, so you can compare cost, speed, and intelligence on the same code. 5.0k 2026-09-17 high

Patterns (5)

Page Summary ~Tokens Updated Conf.
Composite scoring Break a ranking judgment into independent Score dimensions, normalize each to 0–1, and combine them with weights your code owns and can retune. 2.1k 2026-09-17 high
Confidence-gated routing Treat confidence as a second decision axis: one floor below which nothing is automated, then a per-action threshold sized to the consequences of being wrong. 2.2k 2026-09-17 high
Speculative fan-out Put every question your decision tree could need into one call, including speculative ones, and let code discard the irrelevant answers. 2.2k 2026-09-17 high
Intent routing Classify the request with one cheap Jev call, then invoke the cheapest adequate handler: deterministic code, a specialist LLM, or a human. 2.6k 2026-09-17 high
Patterns overview Catalog of TypeSafe's four architectural patterns — fan-out, confidence routing, composite scoring, intent routing — and when to reach for each. 1.8k 2026-09-17 high

Cookbooks (19)

Page Summary ~Tokens Updated Conf.
Cookbook: Structure recovery (autoformat) Rebuild Markdown from de-formatted plain text in two requests: Nouls stitch hard-wrapped lines, then Choices classify each block; code renders every character from the input. 4.5k 2026-09-17 high
Cookbook: Autoresearch feature discovery An LLM proposes Jev questions, Jev answers them per row to build numeric columns, CatBoost trains and reports back: 38 questions after five rounds reach 1.77 held-out RMSE on wine scores. 4.7k 2026-09-17 high
Cookbook: Double-checking citations One Choice question decides whether a quote's surrounding section supports, contradicts, or says nothing about an LLM's claim; a string match catches fabricated quotes first. 2.8k 2026-09-17 high
Cookbook: Classification using confidence Classify SEC filings into 75 SIC industry groups with one Choice; when confidence < 0.9, report the broader division instead — 40% correct becomes 70%, at no extra call. 2.9k 2026-09-17 high
Cookbook: Classifying RAG passages Four Noul questions per retrieved query-passage pair, routed by ordered thresholds, decide what reaches the generator as evidence, as conflict, or not at all. 3.6k 2026-09-17 high
Cookbook: Self-consistency — choices Runs an 8-Choice moderation rubric 15 times across Jev and six LLM conditions; a 0.60 top-probability floor lifts Jev's decision agreement from 90.8% to 99.2% at 74.2% automation. 4.3k 2026-09-17 high
Cookbook: Self-consistency — nouls Runs a 14-Noul claims rubric 15 times against Jev and six LLM conditions; Jev's mean probability std dev is 0.0102 at 111ms per call, and an uncertain band routes 0.30-0.70 to a human. 3.7k 2026-09-17 high
Cookbook: Date extraction Seven Choice questions read a date's parts off a document in one call; code does all the calendar math and gates on the weakest part's confidence. 3.9k 2026-09-17 high
Cookbook: Knowledge graph entity alignment One three-level Score decides merge / curate / leave-unlinked for 450 candidate entity pairs, with three Nouls riding along to tell the curator which field disagrees. 3.0k 2026-09-17 high
Cookbook: Function calling Map typed Python functions to Jev Choice/Noul questions so a natural-language command becomes a validated call with a per-argument confidence. 2.9k 2026-09-17 high
Cookbook: Hierarchical classification Walk a deep taxonomy to a leaf by asking one Choice per sibling set, and keep K paths alive with a geometric-mean beam search instead of a greedy walk. 2.9k 2026-09-17 high
Cookbook: Guardrails for LLMs Screen every LLM input and output with one Jev request: four hazard Nouls plus a severity Score, then route pass / review / block / support on thresholds you own. 3.1k 2026-09-17 high
Cookbooks overview Catalog of all 18 TypeSafe cookbooks — goal, primitives and domain each — plus how a cookbook is structured and why the published pip line does not work. 3.4k 2026-09-17 high
Cookbook: Parallel questions Batching 13 questions about one 54k-character document into a single Jev call was 12.2x cheaper and 10.0x faster than 13 single-question calls, with identical answers. 3.1k 2026-09-17 high
Cookbook: Pre-parsed value extraction A recall-tuned regex finds candidate spans, a Jev Choice over those exact spans picks the one the question asks for, and code copies it verbatim and normalizes it. 2.7k 2026-09-17 high
Cookbook: Re-ranking BM25 builds a 30-candidate shortlist per CLERC query, then one Noul per query-candidate pair re-sorts it: top-1 rises 5% to 18%, top-10 38% to 62% for $0.0645. 2.6k 2026-09-17 high
Cookbook: SDE cascade Extract with a cheap model, verify every field with a battery of Jev Nouls framed so true means wrong, and escalate to a reasoning model only when a flag fires. 4.0k 2026-09-17 high
Cookbook: Line-by-line search Tag a document's lines with ids, rank them with one Choice over those ids, and pair it with a Noul that says whether the document answers the query at all. 2.4k 2026-09-17 high
Cookbook: Skill suggestion Two Jev requests per agent turn — rank 182 skills then re-read the top three — cut wrong skill loads from 16.8% to 7.3% and needless loads from 9.8% to 4.0%. 3.8k 2026-09-17 high

Syntheses (4)

Page Summary ~Tokens Updated Conf.
FAQ for agents and developers Forty answers about Jev — what it is, the API, the SDKs, design, confidence, limits and pricing, company and legal — each linked to the page carrying the detail. 4.9k 2026-09-17 high
Glossary Sixty-nine Jev, TypeSafe, API, SDK and evaluation terms defined in one or two sentences, alphabetical, each linked to the page with the full contract. 4.7k 2026-09-17 high
Jev vs LLM JSON mode / structured outputs What Jev gives you that an LLM in JSON mode does not (and vice versa), a row-by-row comparison, when to pick each, how to combine them, and how to benchmark the pair honestly. 4.3k 2026-09-17 medium
Versions and timeline (models, SDKs, API, company) Every datable fact in the wiki on one axis: company, model versions, API, both SDKs, the plugin and adapter, legal documents, repo commits, and where the dates disagree. 4.0k 2026-09-17 medium

Entities (11)

Page Summary ~Tokens Updated Conf.
Blog: Lies, Damned Lies, and Benchmarks 2026-09-11 post against benchmaxxing, ending in TypeSafe's commitment to publish no standard benchmark table and to retire dated eval snapshots. 1.7k 2026-09-17 high
Blog: The Bitterest Lesson 2026-09-10 post arguing the ML priority order is doing the right task > data > compute > algorithms, using InstructGPT vs GPT-3 as the proof. 1.4k 2026-09-17 high
Blog: Introducing System One Models & Jev (2026-09-15) Digest of TypeSafe's launch post: System One vs LLM comparison table, price/speed claims, the evidence section, and the caveats the post states itself. 2.6k 2026-09-17 high
Blog: AI: too good to be true, too bad to be useful (2026-06-19) Stub: the body of this 2026-06-19 TypeSafe post was not captured (JS-rendered); only title, date, category, and the homepage teaser are known. 0.7k 2026-09-17 low
typesafe-ai GitHub organisation and repos The github.com/typesafe-ai org: two SDKs, the agent skill marketplace, the System One adapter, plus infra repos and two forks. 2.3k 2026-09-20 high
Jev (model) Jev is TypeSafe's first System One model: text in, typed decisions with calibrated probabilities out, $0.042/MTok input, 64k context, 70-500 ms. 1.6k 2026-09-17 high
Manifesto: Composable AI — Build Prod, Not God TypeSafe's manifesto argues the bottleneck is composability, not intelligence, and lays out a three-step plan to make AI a dependable software primitive. 1.7k 2026-09-17 high
Press and third-party coverage Outlet-by-outlet digest of Jev's launch coverage (2026-09-15/16), separating TypeSafe's claims from each outlet's own assertions and caveats. 3.5k 2026-09-17 medium
Founders and team TypeSafe AI's three founders (Diogo Almeida CEO, Sasha Sheng COO, Erik Gafni CTO), the team's stated culture and values, and the SDK maintainers. 1.6k 2026-09-21 high
TypeSafe AI (company) TypeSafe AI, Inc. is a San Francisco frontier AI lab that left stealth on 2026-09-15 with $40M led by DCVC and Jev, its first System One model. 2.5k 2026-09-17 high
console.typesafe.ai (console + playground) console.typesafe.ai is TypeSafe's Web Interface: log in with Google or an email code, try Jev in the Playground, and mint API keys. 1.1k 2026-09-17 medium

Tags