Explainers
What Jev is, what System One means, and how the model is meant to be used.
What is Jev?
Jev is TypeSafe AI's flagship System One model: unstructured state in, typed probabilistic decisions out. No text generation.
What is a System One model?
System One models make fast, structured decisions for software. Jev is TypeSafe's flagship and first model in that class.
RLCD: Reinforcement Learning for Calibrated Decisions
TypeSafe trains Jev with RLCD — reinforcement learning aimed at epistemically honest probabilities on System One tasks.
Jev vs an LLM: the official contrast
TypeSafe trains Jev for calibrated decisions, not generated text. Here is the contrast they publish — and the caveats they attach.
Why Jev gives up string generation
TypeSafe argues that dropping free-text output buys type safety, parallel answers, and software-shaped use.
Jev System One versus System Two reasoning | TypeSafe Jev System One
Compare fast decisions with deliberation — System One and System Two decisions. Independent unofficial notes; confirm against docs.typesafe.ai.
Jev 1.13: the current public model
jev-1.13.0 is the versioned ID behind jev-latest and jev-preview as of the public models page. Pin it if you tune thresholds.
State: the material Jev evaluates
State is a string, object, or array of text. All questions in a request see the same state. Images and audio are not supported.
State and memory across Jev decisions | TypeSafe Jev System One
Understand stateful decision workflows — Jev state and memory. Independent unofficial notes; confirm against docs.typesafe.ai.
Atomic questions, composed in code
Ask one snap judgment per question. Split multi-factor work. Combine answers with weights you own.
TypeSafe Jev System One decision model overview | TypeSafe Jev System One
Understand how the model works — TypeSafe Jev decision model. Independent unofficial notes; confirm against docs.typesafe.ai.
Signals and actions in a Jev decision | TypeSafe Jev System One
Understand how signals become actions — Jev signals and actions. Independent unofficial notes; confirm against docs.typesafe.ai.
Evidence-first decision design for Jev | TypeSafe Jev System One
Learn how evidence supports decisions — evidence-first Jev. Independent unofficial notes; confirm against docs.typesafe.ai.
Policy-aware decisions with Jev
Hard policy stays in code. Jev judges language-y gray zones. Confidence is not an allowlist.
Confidence scores and escalation in Jev | TypeSafe Jev System One
Learn when a decision should escalate — Jev confidence escalation. Independent unofficial notes; confirm against docs.typesafe.ai.
Uncertainty handling in the Jev model | TypeSafe Jev System One
Learn how to represent uncertainty — Jev uncertainty handling. Independent unofficial notes; confirm against docs.typesafe.ai.
Safe defaults in TypeSafe Jev | TypeSafe Jev System One
Design conservative fallback actions — Jev safe defaults. Independent unofficial notes; confirm against docs.typesafe.ai.
Jev decision boundaries
How Jev’s closed label set, confidence, and code thresholds draw the decision boundary — and where a trained classifier still wins.
Decision ownership and accountability in Jev | TypeSafe Jev System One
Clarify ownership for automated decisions — Jev decision ownership. Independent unofficial notes; confirm against docs.typesafe.ai.
Human-in-the-loop Jev workflows | TypeSafe Jev System One
Understand human review handoffs — Jev human in the loop. Independent unofficial notes; confirm against docs.typesafe.ai.
Deterministic outputs for Jev decisions
Jev closes the answer set; it does not promise bit-identical floats across versions. Pin the model and freeze questions.
Decision traces and auditability in Jev | TypeSafe Jev System One
Learn how to make decisions auditable — Jev decision traces. Independent unofficial notes; confirm against docs.typesafe.ai.
Measuring decision quality with Jev
Measure Jev on policy actions, calibration, and review rate — not an invented leaderboard or vendor homepage multiples.
Jev offline evaluation
Run your own Jev eval harness: labeled tickets, frozen questions, pinned jev-1.13.0. Do not invent 193x numbers.
Production feedback loops for Jev | TypeSafe Jev System One
Improve decisions with observed outcomes — Jev production feedback. Independent unofficial notes; confirm against docs.typesafe.ai.
Jev decision latency
What TypeSafe’s 70ms–500ms and 40x–200x (and 193.6x) claims measure — and how to time your own path. No invented numbers.
Context selection for Jev decisions | TypeSafe Jev System One
Understand which context a decision needs — Jev context selection. Independent unofficial notes; confirm against docs.typesafe.ai.
Model-agnostic Jev decision design | TypeSafe Jev System One
Understand portability across models — model-agnostic Jev. Independent unofficial notes; confirm against docs.typesafe.ai.
Jev data handling and customization
TypeSafe says Jev is not trained on customer requests and is not LoRA-adapted per account. ZDR is a legal topic.
Jev language support
English is Jev's primary training language. Other languages are accepted with currently lower accuracy.