# Limitations

- One CL-Bench task out of six (blind spectrum monitoring). The CAIN arms ran 3 runs; the leaderboard uses 5; plain ICL ran 1 run.
- Two models: claude-opus-5-5 (all arms) and claude-haiku-5-5 (basic retrieval only; its other arms stopped when the API credit ran out). Leaderboard rows are other model versions, for context only, not a controlled comparison.
- Same-project run; the signing key is self-generated; 0 independent witnesses.
- On BSM the per-scan reward is hidden from the system, so skill promotion and causal retrieval are not evaluable on this task.
- The 6-hour mission runs on one host with SQLite: crash-fault tolerance, not Byzantine, not multi-host; the partition is heartbeat suppression; tasks are deterministic digests, not inference.
- MemoryArena, AMA-Bench and AgentLAB were not run (judge model / budget).
- VAGI infrastructure is the 2027 goal and is not demonstrated.
