docs/evaluation.md

# Evaluation

Evaluation is computed from simulator rollouts, perturbation suites, baseline comparisons, and post-save inference checks.

## Run

```bash
.venv/bin/python scripts/evaluate_baselines.py
.venv/bin/python scripts/evaluate_all.py
.venv/bin/python scripts/evaluate_compare_runs.py --baseline outputs/reports/baselines.json --candidate outputs/reports/benchmark_report.json --output outputs/reports/improvement_report.json
```

## Main Artifacts

- `outputs/reports/benchmark_report.json`
- `outputs/reports/baselines.json`
- `outputs/reports/grpo_ablation_report.json`
- `outputs/reports/improvement_report.json`
- `outputs/plots/*.png`
- tracked mirrors under `docs/results/`

## Metric Families

- Offline policy quality: `avg_reward`, `legal_rate`, `success_rate`.
- Robustness under perturbations.
- Dosing-specific target attainment and toxicity avoidance.
- Calibration and abstention.
- Process fidelity and invalid-action behavior.
- Subgroup and explainability summaries.
- Failure visibility and anti-cheat counts.

## Improvement Gate

Final comparison must show positive or non-regressing behavior on:

- average reward
- legality rate
- success or justified-safe-resolution rate
- process fidelity
- timeout rate
- failure visibility

Current tracked smoke artifacts are not final evidence: `docs/results/improvement_report.json` currently records `improved: false`. Replace it after real SFT/GRPO training.