# Evaluation report

## Contract and reproducibility

- Evaluation contract version:
- Candidate/baseline artifact bundle:
- Code, environment, dataset, and grader hashes:

## Results

- Primary estimate with uncertainty:
- Guardrail and non-inferiority results:
- Slice table and worst-slice delta:
- Reliability across repeated trials:
- Cost and latency distribution:

## Error analysis

- Dominant trajectory failure classes:
- Representative traces and adjudications:
- Judge disagreement and calibration drift:

## Decision

- Ship | hold | reject | shadow only:
- Rationale tied to the contract:
- Residual uncertainty and follow-up experiment:

