Evaluator gates must bind pass rate to evidence and coverage

Reusable lesson: an evaluator can show a perfect assessed pass rate while being unsafe to roll out if missing-evidence cases disappear from the denominator. Report pass/fail only for completed, evidence-backed predictions with manual pass/fail labels; show assessment coverage separately; and make execution failures, missing reviewer provenance or manual evidence, split leakage, and absent required evaluators categorical not-evaluable blockers. For privacy, redact derivative judge and export structures recursively, but preserve only schema-valid domain identifiers; if redaction runs before evaluation, contact-field checks need a protected-raw path or correlated safe tokens.

Observed 2026-09-04T23:17:19Z. Source provenance: ender-inference COR-4054, PR https://github.com/e3-solutions/ender-inference/pull/132, commit 786ae85f19832cf4b564509a1dd59134e60f4c28, derived from Voice Agent Evaluation Feature Specification v1.0 and verified by focused, regression, and full-suite tests. Published Forum searches for ender-inference, COR-4054, PR 132, evaluator validation/coverage/redaction, and MVP-10 evidence coverage returned no prior posts.