Primary sources

Readings

The 101 is a beginner rewrite. These are the pages it is allowed to lean on. Prefer the original when you need exact wording.

Official launch

Introducing System One Models and Jev ↗

Sep 15, 2026

Diogo Almeida’s founding essay: new architecture, parallel sampler, RLCD. Pricing, latency band, naming (Kahneman / Jevons), and the workflow-eval caveat on homepage multiples.

Manifesto

Composable AI: Build Prod, Not God ↗

Fetched Sep 20, 2026

The bottleneck isn’t raw intelligence — today’s intelligence is hard to build on. AI as a primitive for semantic judgment; code for exact computation.

Official thoughts

Lies, Damned Lies, and Benchmarks ↗

Sep 11, 2026

Anti-benchmaxxing: no standard benchmark table in model releases. Dated snapshots, then retire them. Publish caveats and evidence that looks bad.

Official marketing

TypeSafe homepage ↗

Fetched Sep 20, 2026

“We took the opposite research direction / not chat.” Decisions, calibrated confidence, more like code. Also lists The Bitterest Lesson (deep URL not confirmed here).

Official evals

Workflow evals ↗

Fetched Sep 20, 2026

Four published workflows (security incidents, agent traces, invoices, customer service). Source of 193.6× / 444.6×. “Structure is always better.”

Official docs

How to build with System One ↗

Fetched Sep 20, 2026

AI-powered software, not agents. Code owns control flow. Most queries ~100 ms (docs claim).

Official docs

Jev 1.13 jaggedness ↗

Reviewed Sep 17, 2026

Honest weak spots: not trained to generate text; not a calculator; counting, dates, indirection, adversarial state.

Video talk (YouTube embed)

AI: too good to be true, too bad to be useful ↗

Jun 19, 2026 · AI Council

The typesafe.ai blog URL is not a prose essay — it embeds Diogo Almeida’s AI Council talk. Watch on YouTube. Page: typesafe.ai/blog/ai-too-good-to-be-true-too-bad-to-be-useful-typesafe-ai

Third-party · LangChain

Building a Harness with Jev ↗

Sep 17, 2026

TypeSafeClassifier, ModelRouterMiddleware, AutoModeMiddleware. Company 200×/400× restated as TypeSafe reports. Community names (Browserbase, trading, email triage) are LangChain’s, not this 101’s.

Third-party · Langfuse

Using TypeSafe’s Jev for evals ↗

Sep 18, 2026

Atomic eval questions; Noul has no separate confidence field; context rot; cannot abstain. Relays Good Start Labs 6,003-check bench — third-party numbers.

Platform changelog

Jev on Vercel AI Gateway ↗

Sep 17, 2026

AI SDK 7 evaluate API, model typesafe-ai/jev. Example: queue + priority + refund-review. Gateway options ≠ TypeSafe defaults.

Platform docs

Jev on Cloudflare Workers AI ↗

Fetched Sep 20, 2026

Model typesafe/jev. Worked examples: support routing, structured refund review, account-risk Score + escalate Noul.

Third-party teaching map

Jev use cases (AY Automate) ↗

Sep 19, 2026

Seven-pattern rewrite of official buckets plus confidence-gated flow. Author states they have not used Jev on client work.

Community index

Awesome Jev ↗

Refreshed Sep 17, 2026

Repos, SDKs, MCP servers, routers, and demos. Star counts and project blurbs are the index author’s. Useful map, not an official catalog.