# DriftShield — eval snapshot (`eval_compare.md`)

This file is generated by `scripts/generate_proof_artifacts.py` for hackathon / README evidence. **Baseline** = two-step hand rollouts per task (opens case then submits a weak resolution). **Trained** column uses the deterministic scripted strong policy in `inference.fallback_action` (same trajectories as `tests/test_driftshield.py`), standing in for a GRPO-tuned LoRA until you attach a real adapter under `outputs/`.

---

# Baseline vs Trained — DriftShield

- Base model: `Qwen/Qwen3-1.7B`
- Adapter:    `inference.fallback_action (scripted strong baseline, not LoRA)`
- Curriculum: `driftshield`
- Episodes:   1 per run per task

## Component means (higher is better except penalty)

| Metric | Baseline | Trained | Δ |
|--------|----------|---------|----|
| Total (progress_score) | +0.347 | +0.962 | ▲ +0.615 |
| Investigation | +0.050 | +1.000 | ▲ +0.950 |
| Routing | +0.750 | +1.000 | ▲ +0.250 |
| Reply quality | +0.000 | +1.000 | ▲ +1.000 |
| Groundedness | +0.750 | +1.000 | ▲ +0.250 |
| Submission | +0.750 | +1.000 | ▲ +0.250 |
| Pass rate (total≥0.5) | +0.000 | +1.000 | ▲ +1.000 |
| Done rate | +1.000 | +1.000 | · +0.000 |
| Penalty total (↓) | +0.280 | +0.037 | ▼ -0.243 |
| Turns (mean) | +2.500 | +4.250 | ▲ +1.750 |