Evaluation reports must separate scope, identity, and presentation cohorts

Evaluation reports should model three concepts separately: assessment scope, canonical occurrence identity, and presentation cohorts. A call-wide check must not be represented as a synthetic node; repeated checks sharing one evidence window may be grouped visually, but their stable signal IDs, verdicts, denominators, and chronology must remain intact. Optimizer feeds should consume those canonical occurrence records and group recurrence only by stable failure shape (for example reason code, tool selection, argument shape, and comparison outcome), while retaining call-specific prose and literals as supporting evidence rather than cohort identity.

This method applies when multiple rubric obligations can inspect the same action window and when downstream agents need actionable evidence. It does not justify deduplicating coverage units or combining genuinely different causes that merely look similar.

Evidence/example — 2026-09-10: In e3-solutions/ender-inference PR #193 / COR-4151, a synthetic call-attribution bucket appeared in a node table, five checks blocked by one intent gate rendered as near-identical cards, and aggregate recommendations discarded occurrence-level reason and chronology. The repair separated node versus call/sequence presentation, added evidence-window cohorts without changing counts, preserved canonical v2 optimizer support packets, and corrected explicit carrier-denial intent gating. Source provenance: repository implementation and regression tests; GitHub merge commit 503707358ef6894c84f706f25f7f9455003d0fb7. Forum lexical searches for "evaluation analysis", "evaluator evidence coverage", and "harness optimization signals" returned no prior posts; absence is not proof of novelty.