Jev as a second-perspective judge

Jev returns typed judgments (Choice, Score, Noul) with probabilities; it does not reason or generate text. Code owns retrieval, arithmetic and policy; Jev supplies one narrow judgment per question. Read the typesafe-ai skill and the live docs (https://docs.typesafe.ai/llms.txt) before changing request shapes.

Client

python3 scripts/jev.py REQUEST.json [OUT.json] (POST /v1/systemone, model defaults to jev-latest) and python3 scripts/jev.py --models. The key comes from $TYPESAFE_API_KEY or ~/.config/typesafe/api-key (mode 600); never put it in a request file, artifact, skill or command line. Billing is per input token; output is free.

Pattern 1: lever triage (composite scoring)

State = the task, the rules (gates, time guard, editable files, cost weights) and the candidates with their ledgers. Per candidate ask two independent Scores (feasibility under the rules; share of the required gap) and one Choice over all candidates for "the next hour of work". Multiply in code. Treat it as a tie-breaker and a blind-spot detector: a candidate it rates plausible but unmeasured is the next thing to measure, cheaply. Do not trust the rating itself.

Pattern 2: citation check of a public draft

Code splits the draft into claim blocks (those with job ids or large numbers), retrieves the log entries that share the most ids/numbers, and asks one Choice per (claim, entry): supports / contradicts / says_nothing. Only confidence >= 0.8 stands on its own; route every "contradicts" to a human check. In practice most contradictions came from retrieving a neighbouring job's entry; retrieve by the claim's own id first. It still caught an actual overclaim ("twelve records" when the board's twelve include another participant's starting record).

Limits

Jev judges only the state you give it; it cannot verify soundness, run code, or invent constructions. Never let a Jev answer authorize a submission, a post, or spend: those stay with the measured gates and the user.