Score free-form evidence arguments by grounded trigger coverage, not one arbitrary gold span.

Score free-form evidence arguments by grounded trigger coverage, not one arbitrary gold span.

General lesson: When a tool argument must quote literal source evidence, exact equality to a single annotator-selected string creates false negatives. Freeze the minimal affirmative trigger as source offsets while outputs are hidden, then accept any candidate evidence that is (1) a nonempty exact source substring and (2) has an occurrence whose interval fully contains an approved trigger interval. Keep tool identity, closed arguments, enums, and call count exact. Negative, quoted, reported, conditional, or merely discussed trigger words receive no affirmative interval, so literal-but-invalid evidence still fails.

Applicability and limits: This works for evidence-bearing tool calls where multiple literal spans can faithfully support the same decision. It does not resolve route ambiguity, permit paraphrases, or replace intent annotation. Mixed-route precedence conflicts must be marked ambiguous or settled by a separately versioned policy before evidence scoring.

Evidence/example (2026-09-16): In a synthetic end-call scorer design, both “Goodbye” and “Goodbye.” validly cover the same affirmative trigger, while “Good”, hallucinated text, or “hang up” inside “do not hang up” fail. No model outputs or real calls were used.