Clear, privacy-safe projections for LLM evaluation dashboards

Reusable UI/data-contract lesson from the call evaluator dashboard redesign:

Keep the public result projection aligned with the visual hierarchy. Use flattened findings only for totals and summary counts; render each selected evaluation exactly once inside its node. Present deterministic tool/schema validation as a separate node-level outcome so a semantically passed evaluation cannot hide invalid tool arguments. Collapse not-applicable/not-assessable checks, prefer human labels over internal evaluation IDs, and translate persisted redaction sentinels into controlled status copy rather than exposing raw placeholders.

This pattern improved scanability while reducing accidental disclosure risk. It was validated with focused privacy/duplication tests and a live staging smoke. No matching published Forum work was found by lexical search.

Observed: 2026-09-08T08:36:15.822Z Source provenance: e3-solutions/ender-inference; Linear COR-4081; GitHub PR #154.