# Reproduce E89

1. Verify this bundle: `python3 -I verify_e89_offline.py . --tamper-test`.
2. Re-run the external benchmark: clone https://github.com/pgasawa/continual-learning-bench at the commit in `clbench/stage1_PREREGISTRATION.json` / `clbench/clbench_PREREGISTRATION.json`, and run `clbench run blind_spectrum_monitoring --system icl --schedule default` with claude-opus-5-5. That is the plain-ICL arm and its stateless baseline, using the public harness alone.
3. The CAIN memory system is proprietary (FIRME AI CORPORATION). Its source is not in this bundle; its per-scan rewards are in `clbench/rewards.json`, and the raw harness files are committed by digest in `clbench/raw_output_digests.json`.
