DGXAI/driftcall-env

1# Cell 19 — Final Evaluation (Post-Training LoRA)2 3`eval_final(checkpoint, ..., baseline=baseline_report)` runs the trained LoRA4on the **same** 50 paired episodes used by the baseline (evaluation.md §3.1)5and stores the paired-difference 95% CIs under6`EvalReport.breakdown['paired_ci']`.7 8**Contract:** evaluation.md §2.1, §3.1, §3.3, §3.8, §5 `EpisodeSetLeakError`.9 10- `EpisodeSetLeakError` raised at entry AND exit if `baseline.episode_ids ≠11  val/briefs.jsonl[0:50]` or the post-rollout report's IDs diverge.12- Paired bootstrap CI seed = `20260428` (evaluation.md §2.4).13- Wall-clock budget 20 min — same ceiling as baseline.14