Evaluation Charter and Specification — Document Frontier

Origins: research/evaluation/00-charter.md; 02-evaluation-specification.md

Knowledge and limitations

These artifacts define verified repository outcomes, outcome classes, integrity audits, layered graders, run metadata, and containment. Thresholds and constructs are principled but not externally calibrated.

Five highest-value opportunities

1. RFR-001: Independent construct audit (“Task Integrity Audit”).
2. RFR-003: Estimate grader error (“Grader Stack”).
3. RFR-002: Observe variance under the run protocol.
4. RFR-006: Validate minimum metadata against economic decisions.
5. RFR-008: Test the containment requirement.

Challenge: Declared outcome classes may omit ambiguous mixed outcomes; natural-run disagreement analysis should refine rather than assume completeness.