Evaluation Results and Cycles — Document Frontier

Origins: research/evaluation/07-results.md; cycles/2026-07-23-cycle-001.md; cycles/2026-07-28-cycle-002.md; cycles/2026-07-28-cycle-003.md

Knowledge and limitations

Two contracts validate, seven designer probes classify as intended, and blind export detects no declared exposure. No agent was evaluated; review is not independent; exposure checks are narrow.

Five highest-value opportunities

1. RFR-001: Independently review both pilots.
2. RFR-002: Generate natural repeated outcomes.
3. RFR-003: Compare probe behavior with expert labels.
4. RFR-008: Audit fixture isolation and containment.
5. RFR-004: Expand beyond two documentation-oriented pilots.

Challenge: Seven-of-seven is not an uncertainty estimate from independent samples; it only establishes implementation consistency with designed fixtures.