Reference
Resources & next steps
Official links first. Secondary press is labeled. Quotes are exact and dated. This page is a map, not a substitute for TypeSafe’s docs.
The 101 path
- 01 · What is Jev? — A System One model: typed decisions software can use, not chat replies.
- 02 · How Jev works — State in. Choice / Score / Noul questions in. Typed answers out.
- 03 · Use cases — Triage, gates, verify, rerank, real-time control — forty decision slots.
- 04 · Confidence & thresholds — When to auto-act, escalate to a human, or call a bigger model.
- 05 · First call mindset — Access, one question, one threshold. Pin official docs for the SDK.
- 06 · FAQ — Access, primitives, claims, limits, and naming — from TypeSafe materials.
Also:Updates ·Ecosystem ·Talks ·Social ·Readings ·What’s the difference?
Official TypeSafe
Official site / waitlist ↗
Primary product home. Join Waitlist CTA as of this guide’s compile date.
Launch post: Introducing System One Models & Jev ↗
Sep 15, 2026. Richest primary source for positioning, pricing, latency, naming.
Manifesto: Build Prod, Not God ↗
Composable AI. Bottleneck is buildability, not raw intelligence.
Team ↗
Founders and office. Confirm bios before quoting.
Docs index ↗
Concepts, API, SDKs, cookbooks. Prefer this over frozen samples.
Quickstart ↗
Endpoint, keys, Python SDK.
Models ↗
Price, limits, context, language caveats, pin jev-1.13.0.
How to build with System One ↗
AI-powered software, not agents. ~100 ms typical (docs).
Use-case map ↗
Official brainstorm by industry and task shape.
Jaggedness (jev-1.13) ↗
Honest weak spots: math, dates, generation, adversarial state.
Workflow evals ↗
Source of 193.6× / 444.6×. Methodology lives here.
Anti-benchmaxxing ↗
Why TypeSafe refuses standard public bench tables.
AI: too good to be true (YouTube talk) ↗
Video, not a prose essay. The typesafe.ai blog URL only embeds this AI Council talk: youtube.com/watch?v=o-y1HJ6buGQ
Secondary / disambiguation
TechCrunch (Sep 18, 2026) ↗
Secondary press. Developer-reaction quotes. Not TypeSafe metrics.
SecondaryVercel AI Gateway changelog ↗
Platform path. Do not equate Gateway features with TypeSafe defaults.
SecondaryCloudflare Workers AI · typesafe/jev ↗
Platform docs with routing, refund, and risk examples. Not TypeSafe’s own site.
SecondaryLangChain: Building a Harness with Jev ↗
TypeSafeClassifier + experimental routers/guards. Company multiples restated as TypeSafe reports.
SecondaryLangfuse: Jev for evals ↗
Eval-scoring recipe. Relays a third-party rubric bench. Noul has no separate confidence field.
SecondaryAwesome Jev (community index) ↗
Repos and demos collected by a third party. Star counts are theirs.
SecondaryCyera SafeType (different product) ↗
Disambiguation only. ChatGPT privacy extension — not this 101.
Glossary
| Term | Meaning in this 101 |
|---|---|
| TypeSafe AI | San Francisco lab building System One models for machine-native decisions. Not Cyera SafeType. |
| System One Models | TypeSafe’s class of models for fast structured decisions inside software. The name is inspired by Kahneman’s System 1 — a naming metaphor, not a psychology claim. |
| Jev | TypeSafe’s first public System One model. Official launch FAQ: named after William Stanley Jevons / Jevons paradox. |
| RLHF | Reinforcement Learning from Human Feedback — the training paradigm that optimized chat models for human preference. |
| RLCD | Reinforcement Learning for Calibrated Decisions — TypeSafe’s stated training algorithm for calibrated probabilities on System One tasks. |
| State | The content Jev evaluates: a string, JSON object, or array of text. Not a chat transcript of messages. |
| Choice | Pick one option from a set you define. Returns a typed option, a distribution, and confidence. |
| Score | Place the state on an ordered rubric you define. Returns a weighted score plus confidence. |
| Noul | P(yes) for a yes/no question. Official docs: Noul has no separate confidence field — the probability is the uncertainty signal. |
| Confidence | A statistic from the probability shape (Choice / Score) used to gate automation versus escalation. |
| Calibrated | Company goal: across many predictions, stated probabilities match observed frequencies. Calibration is not a guarantee on a single answer. |
| Parallel sampler | Company claim: questions in one request are evaluated independently in parallel, rather than as sequential chat tokens. |
| Speculative fan-out | Ask many questions, including speculative ones, in one call; let code select what to use. |
| Decision slot | Educational term for the place in software where an enumerable judgment is needed — classify, route, score, gate, verify. |
| Cascade | Jev decides → an LLM generates only if needed → Jev optionally verifies → a human if confidence is low. |
| Jaggedness | TypeSafe’s documented weak spots per model version: math, dates, counting, indirection, adversarial state, and no text generation. |
| Workflow evals | TypeSafe’s published harness evaluations versus GPT-6 Astra + Claude Fable 5.1 references. Source of the homepage 193.6× / 444.6× figures. |
| Smart if-statement | Company metaphor: branch on common-sense semantic judgment while code owns control flow. |
| jev-1.13.0 | Current model ID cited in official docs as of this guide’s compile date. Aliases jev-latest and jev-preview can move — pin versions when you tune thresholds. |
| Cyera SafeType | An unrelated ChatGPT privacy browser extension. Different company, different product. This 101 is about TypeSafe AI’s Jev. |
| Vercel AI Gateway | A platform path that lists typesafe-ai/jev and AI SDK evaluate. Gateway promos (including a free window) are Vercel’s, not TypeSafe list price. |
| OpenRouter / Cloudflare / Venice | Hosted ways to call Jev without a new TypeSafe console key. Treat each as a platform: their UI, limits, and pricing can differ from api.typesafe.ai. |
| langchain-typesafe | LangChain’s TypeSafeClassifier plus experimental ModelRouterMiddleware and AutoModeMiddleware. Third-party integration, not an official TypeSafe SDK. |
Sourced quotes used on this site
Never fabricated. Each line has a URL and a fetch or publish date.
Diogo Almeida / TypeSafe launch blog · Sep 15, 2026 · Source“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
TypeSafe AI homepage · fetched Sep 20, 2026 · Source“We took the opposite research direction”
TypeSafe AI homepage · fetched Sep 20, 2026 · Source“We built a new class of models, System One Models, to be natively used by machines. We're building with a new architecture, a new sampler, and a new training algorithm: Reinforcement Learning for Calibrated Decisions (RLCD).”
TypeSafe AI homepage · fetched Sep 20, 2026 · Source“Yet RLHF creates inherent issues such as mode dropping, overconfidence, and lack of reliability. These flaws mean that LLMs require humans-in-the-loop.”
TypeSafe launch blog · Sep 15, 2026 · Source“While Jev gives up string generation, it's optimized for structured outputs and can't hallucinate.”
TypeSafe launch blog · Sep 15, 2026 · Source“No type errors: This would be an easy thing to falsify with just a single counter-example, but it is mathematically impossible.”
TypeSafe launch blog · Sep 15, 2026 · Source“This is where the claims of 193.6x faster, 444.6x cheaper on our home page comes from, and we expect that these are on the higher end of real world gains.”
TypeSafe launch blog · Sep 15, 2026 · Source“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).”
TypeSafe manifesto · fetched Sep 20, 2026 · Source“We're building prod, not God.”
TypeSafe manifesto · fetched Sep 20, 2026 · Source“the bottleneck isn't raw intelligence. It's that today's intelligence is hard to build on.”
TypeSafe manifesto · fetched Sep 20, 2026 · Source“We want AI to work alongside existing software as a primitive that any programmer can invoke for semantic judgement and decisions, while still using code for what it’s best at: exact computation.”
TypeSafe docs · how to build · fetched Sep 20, 2026 · Source“System One is TypeSafe's model for building AI-powered software, not agents.”
TypeSafe docs · jaggedness · reviewed Sep 17, 2026 · Source“jev-1.13 is not trained to generate text.”
TypeSafe docs · jaggedness · Sep 17, 2026 · Source“Jev is not a calculator. We strongly recommend implementing any mathematical logic in code.”
Secondary Armin Ronacher via TechCrunch · Sep 18, 2026 · Source“if this only comes back with 50% probability, maybe this is a coin toss, and I disregard it. But if it's 95%, sure, then I can do something with it.”