DGXAI/driftcall

1# Cell 20 — Reward-Hacking Probe (200 episodes)2 3`probe_reward_hacking(checkpoint, ...)` scans `Rewards.breakdown.anti_hack`4across 200 held-out val episodes (`val/briefs.jsonl[50:250]`) for the 55enumerated exploit classes plus any novel offense codes (threshold = 1).6 7**Contract:** evaluation.md §2.1, §2.3, §3.1, §3.6, §3.8, §4.4, §4.5, §5.8 9- Disjoint from the paired-comparison 50 episodes.10- All 5 known classes always emitted (count == 0 rows kept for the fixed table).11- Novel offense codes surfaced under `ProbeReport.novel_classes` and flagged12  with `UNKNOWN EXPLOIT CLASS` in the markdown writeup.13- `ProbeOnBaseModelError` raised if `model_path == "base"`.14- `ProbeInsufficientSamplesError` raised if `episodes < 50`.15- Wall-clock budget 60 min — `EvalBudgetExceededError` on overrun.16- No LLM-as-judge anywhere in the scoring path.17