Evaluate decisions against evidence

Use the current TypeSafe API and confidence guide when implementing requests.

Define the claim and candidate passages; keep expected labels outside model state. Use a Noul for whether a passage fully supports a claim, or a Choice when code must distinguish support, contradiction and silence. Include entity, time, qualifiers and scope. Missing evidence is not a denial. Copy receipts from source spans in code; a verbatim receipt establishes provenance, not correctness. Measure retrieval coverage separately from the judgments on retrieved passages.

Include ordinary cases, missing facts, contradictions, negation, different entities, numerical mismatches, dropped conditions, changes from allowed to required, valid inferences and instructions embedded in documents. Keep synthetic fixtures public and customer inputs private.

Sweep thresholds on development cases and freeze the policy before testing unseen cases. Report accuracy on accepted cases together with review rate, coverage and errors per category. Use balanced accuracy for uneven classes. Neither confidence nor a chosen-label probability is an accuracy guarantee, and repeated calls do not increase the number of independent cases.