Named for Kahneman’s fast, automatic judgment. The caller declares the legal answers in the request; the model returns a distribution over exactly those. Jev is TypeSafe AI’s first public System One model.

State is text-only and English-primary. Labelled structure is preferred over raw blobs because field names carry signal; images, audio and video must be serialised to text first. Budget: 32k tokens with the longest question inside a 64k-token request.

Each question gets a caller-chosen key, an instructions string, and primitive-specific fields (options for Choice, criteria for Score). One focused judgment per question — fused judgments smear probability mass.

Answer fields: the winning choice, per-option probabilities, and confidence. Always declare an "other" option — without an escape hatch, off-menu inputs are force-fit into the nearest declared option.

The declared criteria are ordered rungs; the returned score is the expectation over their probabilities and can land between rungs. Use it when the judgment is "how much / how severe", not "which one".

Returns a bare probability with no separate confidence field: for two outcomes the spread is already visible. The natural input to a threshold policy (act above τ, escalate below).

An aggregate guarantee, not a per-case one: it makes thresholds into policies with a measurable error budget. It does not survive distribution shift and does not promise low error rates on hard tasks.

Concentrated distributions give high confidence, flat ones low. Computed deterministically from probabilities, so it can never disagree with them; absent on Noul where the probability already encodes the contest.

Contrasts with RLHF (human preference over responses) and RLVR (verifiable rewards). The optimisation target is the honesty of the reported number, which is what makes thresholds meaningful.

No token-by-token generation, hence no decoder and no prose output. Adding questions widens the pass instead of lengthening it — marginal cost per question is small and outputs are unmetered.

The pattern behind guardrails, routers and stream filters. Choosing the threshold is choosing an error budget; it should be measured on your own distribution because calibration is only promised in aggregate.

Appears in games (per-tick action choice), browser agents (per-page next action), stream filters (per-item keep/drop), and agent guardrails (per-tool-call allow/deny). The model decides; code acts.

All questions evaluate in the same parallel pass, so speculative questions cost scoring, not a round trip. Sequential conditional calls trade away a latency the model never charged.

Confidence routing one level up: the gated object is another model’s output. The SDE-cascade and self-consistency cookbooks run it in both directions.

An honest model near a threshold should wobble; the band turns the wobble into a routed signal. In the nouls cookbook it holds questions like covered and fraud_flag across all 15 repeats.

Keeps state inside the 32k budget and guarantees extracted values are real spans. The canonical instance: BM25 top-30 reranked by Jev, top-1 5%→18% on CLERC.

The shared fix: computation, arithmetic and multi-hop interpretation go in code; the model keeps the judgment. Consistency does not imply identities like p(q) + p(¬q) = 1.

Used in pre-parsed extraction (NONE of the regex spans) and skill suggestion (no skill fits — reject all). The hatch converts force-fit errors into a routed outcome.

The skill-suggestion cookbook’s shape over 182 skills: rank in one pass, then a second pass reads the top three properly and may still return nothing. Wrong-skill loads fell by over half.

Preconditions: you can enumerate the legal answers, including an escape hatch (always declare other). Reads: winning choice + probabilities + confidence. Failure mode without other: off-menu inputs are force-fit into the nearest option.

The answer is a probability-weighted position that can land between rungs, so thresholds act on a continuum. Do not use it for unordered categories: a fake ordering produces a meaningless middle.

Returns one probability and nothing else; the spread is visible in the number itself. The cleanest input to a gate: if (noul > τ) act. Do not mimic Choice with several Nouls — overlapping yeses stop summing to one.

Questions evaluate independently in one pass, so “urgent AND billing?” should be is_urgent plus route-to-billing. Fused questions smear probability mass and hide the contest.

Calibration is promised in aggregate over the training distribution, not on your shifted inputs. Re-derive the threshold when traffic drifts; pinned model versions keep the measurement reproducible.

Answers are unmetered and extra questions are marginal. Within budgets: 64k tok/request, 32k for state + longest question, ~20 rps. Compress or split oversized state upstream rather than truncating silently.

~70–500ms per call makes a ~10 qps loop playable. Where the loop is inside a clock, latency is the feature; for batch streams it is merely convenient.

The LLM reasons freely in prose; only typed decisions act. Seen in jev-guard, jev-axi, pi-verdict, LangChain’s AutoModeMiddleware and model routers.

There is no text output and no chain-of-thought. If the taxonomy itself must be discovered, an open-ended model or a human defines the options first — Jev only judges within a declared space.