TypeSafe Jev
Type: full-code · Vendor: TypeSafe AI (San Francisco) · Language: API (REST, Python, JavaScript/TypeScript) · License: proprietary · Status: active · Status in practice: emerging · First released: 2026-09-15
A decision-only model from the San Francisco lab TypeSafe that answers declared questions with typed choices, scores or yes/no probabilities, so harness code can threshold them instead of parsing generated text.
Description. TypeSafe AI came out of stealth on 15 September 2026 with Jev as its first public model, which it calls a System One model built to make fast decisions inside software rather than to chat. A caller sends a text state plus declared questions of three shapes: Choice (pick one of a fixed set, returned with a distribution and a confidence), Score (a position on caller-defined levels) and Noul (a yes/no probability). Answers cannot fall outside the declared schema, questions over the same state are evaluated independently, and the model is trained for calibration so that its probabilities can be thresholded in code. Most calls complete in about 100 milliseconds. It does not write text, code or explanations, and it is weak at arithmetic and at questions that require cross-step reasoning. SDKs exist for Python and JavaScript, and a Vercel AI SDK provider package exposes it to TypeScript harnesses.
Agent loop shape. No agent loop of its own. The harness calls Jev at a branch point with the current state and a set of declared questions, receives typed answers with probabilities and confidence for each, and applies thresholds in code to pick the next step, send an uncertain case to a human, or fall back. Responses name the versioned model that answered so thresholds can be tuned per model version.
Primary use cases
- branch conditions inside agent harnesses
- intent and ticket routing and triage
- tool and skill selection from a closed set
- safety gates such as classifying shell commands by risk
- checking model outputs or claims against source material
Key concepts
- System One model → semantic-decision-node (docs) — A model built to make fast decisions inside software, not to generate text for people.
- Choice / Score / Noul → structured-output (docs) — The three declared question shapes: pick one option, place on a scale, or give a yes/no probability.
- Calibrated probabilities → semantic-decision-node (docs) — Trained so that answers given 90 percent probability are right about 90 percent of the time, which is what makes code thresholds meaningful.
Patterns this full-code implements —
- ★Semantic Decision Node
Jev is built for exactly this node: declared questions in, typed decisions with probabilities out, thresholds applied by the caller's code.
- ★★Structured Output
Every answer is constrained to the declared Choice, Score or Noul type, so an out-of-schema value cannot be returned.
- ★★Routing
Routing and triage (intent classification, support-ticket routing, model selection) is a documented primary use; benchmarks report strong tool-routing accuracy but near-chance accuracy when asked to…
- ★Cost-Aware Action Delegation
Used as a safety gate that classifies shell commands as read-only, reversible or destructive, feeding a tiered approval policy in code.
- ★★Input/Output Guardrails
Third-party benchmark uses Jev as a prompt-injection detector on tool outputs; low scores were found not to be safe, so the guardrail needs a conservative threshold.
Neighbourhood
Click any neighbour to follow the lineage. Scroll to zoom, drag to pan.
Anti-patterns avoided
Alternatives & relatives
- full-code · framework
Semantic Router
Both return a routing decision for code to act on; Semantic Router uses embedding similarity against example utterances with a score threshold, Jev a trained decision model with calibrated probabilit…
- full-code · framework
Vercel AI SDK
The @ai-sdk/typesafe-ai provider exposes Jev decisions inside AI SDK harnesses.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: needs-verification