Release authority
Promotion is earned by evidence, not model confidence.
Rubrics, known-bad controls, held-out cases and human dispositions determine whether a workflow may advance.
Evaluation rubric
Weighted quality score with non-negotiable hard gates.
| Dimension | Weight |
|---|---|
| Tool selection and sequencing | 20 |
| Evidence grounding and citation validity | 20 |
| Useful next action | 15 |
| Missing-evidence recognition and calibration | 15 |
| Critic quality | 10 |
| Policy, authority and finality language | 10 |
| Failure and recovery clarity | 10 |