Glossary
The Jev ecosystem invents vocabulary at a pace, partly because the model class is new. Every term worth knowing, defined plainly, with links to where each one matters.
The model
System One model
A class of model built for fast, structured decisions inside software rather than conversation. TypeSafe AI introduced the term, inspired by Kahneman's fast System 1 thinking. Optimized for calibrated decisions instead of human preference, and for parallel sampling instead of token generation.
Jev
TypeSafe AI's first System One model, released September 2026. Takes a state plus typed questions and returns typed answers with probabilities and confidence. Named after William Stanley Jevons, the economist who observed that cheaper energy unlocks more uses for it. See the FAQ for the practical version.
Noul
A yes/no question type. You declare a statement, the model returns a single probability from 0 to 1 that the statement is true. Named for "yes or no, likely or not".
Choice
A selection question type. You declare up to 255 named options with short criteria; the model returns the selected option, a probability for every option, and a confidence value. High-cardinality lists above 255 use a two-stage approach: score independently, then choose.
Score
A rating question type. You declare an ordered rubric of 2 to 10 levels; the model returns a probability-weighted score, the distribution over the levels, and a confidence value.
State
The input to a decision. Text or structured data, up to 64k tokens per request. The state is the model's only world: whatever assembles it decides what the model is allowed to know. Filter it in code before sending, because accuracy falls as state fills with material the question does not need.
Confidence
A value reported alongside every answer, computed separately from the answer probability. The architectural use is gating: act automatically above a threshold, send the uncertain band to review. See confidence-gated routing.
Calibration
The property that higher reported confidence means higher accuracy across groups of predictions. It does not guarantee that any individual answer is correct, which is why every serious deployment measures thresholds on its own labeled data. See the rigor page.
RLCD (Reinforcement Learning for Calibrated Decisions)
TypeSafe's training approach for System One models. Unlike RLHF, which optimizes for human preference, and RLVR, which optimizes for programmatically verifiable outputs, RLCD optimizes for calibrated decisions: epistemically honest probabilities on bounded questions.
Parallel sampling
Generating all outputs in a single query instead of one token at a time. This is the source of the two headline economics: 70 to 500 millisecond decisions, and output tokens that cost nothing to meter. It is also why asking ten questions in one request costs roughly the time of one.
Production vocabulary
Confidence gating
The pattern of comparing a probability or confidence value against a threshold where the threshold lives in ordinary code, not in a prompt. Above the gate, the system acts. Below it, the case goes to a human queue or a stronger model. The most universally applied pattern in the ecosystem.
Speculative fan-out
Asking many independent questions about the same state in one request, then keeping only the answers the code needs. Because questions are evaluated in parallel and in isolation, adding questions barely changes response time and never creates context rot between them.
Atomic decomposition
Breaking one compound judgment into several narrow questions and combining the answers in your own formula. A community benchmark measured a four-option Choice at 48.3 percent on a task where four precondition Noul questions reached 98.3 percent.
Abstention
The model cannot abstain on its own: a forced binary picks the least-wrong answer. Abstention-first design means giving every Choice an explicit unknown or none option, and treating unknown as "send to a human". Evidence: with the option available, Jev chose unknown for 95 percent of ambiguous items in a KoBBQ audit.
Context rot
Accuracy loss as the state fills with material the question does not need. The documented remedy is to retrieve and filter in code first, and send only the fields each question requires.
Jaggedness
TypeSafe's published, per-version documentation of known failure modes: literal reading, arithmetic and dates, indirection, context rot, hostile state, forced binaries, option-order sensitivity. Read it before designing an evaluation, not after. The practice of publishing it is itself notable: most vendors do not.
Decision log
The audit trail a System One system needs because the model returns no rationale: state, questions, model version, probabilities, chosen action, and the eventual outcome. If your compliance story requires explanations, the log and your criteria text are where they live.
Calibration drift
Threshold invalidation after a model version changes. The alias jev-latest moves; answers can shift with it. Pin exact versions in production, and rerun your labeled set before adopting a new one.
Access and ecosystem
jev-latest
The moving alias for TypeSafe's current stable model (jev-1.13.0 at the time of writing). Convenient for exploration, dangerous for tuned thresholds. Production integrations should pin the exact version.
Decisions API
The OpenRouter route for Jev (typesafe/jev-1.13). Not a chat-completions endpoint: the request shape carries state and typed questions. One of the two documented waitlist bypasses, alongside Vercel AI Gateway.
Vercel AI Gateway
Hosted access to typesafe-ai/jev through Vercel's AI SDK and gateway, including an experimental evaluate path. No TypeSafe waitlist required. The interface differs from the direct systemOne contract.
Open reimplementation
Community models that reproduce the decision interface rather than the weights: reading option logits from a frozen model, one-pass scoring, constrained decoding. Examples: SemIf, jevlike, Laya, Kev, PocketJev. The latency is competitive; the published accuracy comparisons generally favor the hosted model. See open models.
Last updated 30 September 2026.