A.I.
TypeSafe · System One
Jev
Part of the TypeSafe lab. Concept: System One.
- Status
- Preview, early access
- Released
- 15 Sep 2026 (announced)
- Last checked
- 15 Sep 2026newest source date on this page
Jev is TypeSafe’s first public System One model: a decision model for software, not a chat model for people. You send program state plus typed questions; you get typed answers with probabilities (and confidence on Choice and Score). Early access as of the 15 Sep 2026 announce. Docs default alias is jev-latest (also jev-1.13 in jaggedness notes).
Training
TypeSafe names the training method Reinforcement Learning for Calibrated Decisions (RLCD). The company contrasts RLCD with RLHF (human preference for chat) and RLVR (verifiable rewards). Parameter count, training FLOPs, dataset mix, and knowledge cutoff are not published on the announce or docs pages cited here.
Inference
TypeSafe describes parallel sampling of decisions in one query, not autoregressive token generation. The announce page claims about 70 to 500 ms end-to-end response time for System One shaped queries (vendor timing; West Coast laptop runs noted as nuance). HTTP endpoint is POST /v1/systemone. Input modalities today are text only (strings, JSON objects, arrays of text). Images, audio, and video are not supported yet. Output is typed structured values, not generated prose.
Context
TypeSafe has not published a numeric context window on the docs pages cited here. Company guidance: send only the state the question needs; unrelated detail hurts accuracy (jev-1.13 jaggedness notes).
Methods
Three primitives, mixable in one call, evaluated in parallel against the same state:
| Primitive | Returns (company docs) |
|---|---|
| Choice | choice, probabilities, confidence |
| Score | score, probabilities, confidence |
| Noul | noul (0 to 1) |
Compose judgments in application code (branch, threshold, escalate). Docs say calibration is across groups of predictions and does not guarantee any single answer is correct. Free-text generation is out of scope; use a generative model for that.
Use Jev for atomic decisions your code will act on: classify, route, score, extract, branch. Do not use it to write chat replies, generate code, or invent free text.
Jaggedness on jev-1.13
- Send only the state the question needs. Unrelated detail hurts accuracy.
- Calibration is across groups of predictions. It does not guarantee any single answer.
- Free-text generation is out of scope. Chat replies, generated code, and invented prose are the wrong job.
- These notes are for
jev-1.13. The vendor writeup is in Docs: Jev 1.13 jaggedness.
Benchmarking
TypeSafe’s public stance is that it does not ship a standard public-benchmark table in model releases. New evals are dated snapshots, then retired (antibenchmaxxing post, 11 Sep 2026). Workflow evals on the announce page compare models inside fixed code graphs against a reference average of large chat models (the blog names GPT-6 Astra and Claude Fable 5.1). The homepage markets “193.6x Faster, 444.6x Cheaper” on those System One workflow tasks. That figure is a vendor claim with blog caveats, not an independent third-party scorecard. Independent Arena, HLE, or GPQA-style scores for Jev are not on the pages cited here.
Hardware
TypeSafe has not published training or serving silicon details for Jev on the pages cited here.
Access and pricing
Early access / waitlist. The announce comparison table lists input at $0.042 per million tokens ($42 per billion input tokens on the homepage). Output tokens are described as free or too cheap to meter. The company says long-term pricing sustainability still needs proof. Schema and type-matching: TypeSafe claims type errors are structurally impossible given predefined output shapes. Marketing that says “zero hallucinations” should be read as that claim, not as “never wrong.”