Evaluation model

Reference for how Jev answers map to outcomes. For the short version, see review and fixed classifications.

Bernoulli questions, not true/false ones

A condition asks Jev to estimate p(true) from visible evidence (Jev native type noul, answered {chance: P}); the threshold is the decision boundary on p. Name the visible fact that moves p in guidance ("count only ...") rather than restating truth conditions.

Confidence vs probability

Classification confidence and policy severity stay separate. Thresholds gate on probability (P) only; confidence is displayed (for example P=0.86 confidence=0.73) but never gates, suppresses, or applies an outcome. An important gate does not block unless its probability crosses the configured gate threshold; a high-confidence advisory remains non-blocking.

Which threshold governs what

gateThreshold governs gate conditions, secrets, AND fix_now-mapped choice labels; advisoryThreshold governs advisory conditions and advisory-mapped labels (a fix_now label between the two reports an uncertain-gate note).

Below-cutoff fallthrough

A choice whose top label scores below its outcome's cutoff falls through to approve for that scope — no failure, no note, no fallback to the next label — while the classification line is still printed.