# E91 Research → Implementation Matrix

| Finding (source) | E91 implementation | Test / evidence | Owner |
|---|---|---|---|
| Naive ICL ≥ memory systems (CL-Bench) | arm B (recent experience in context) and arm C (BM25) vs A, D and G | H2 in `benchmarks/e91/EXPERIMENT_PROTOCOL.md`; `rows_heldout_{A,B,C}.jsonl` | ec |
| Similarity retrieval loses causal and objective information (AMA-Bench) | G keys skills on task signature plus verified condition ranges, not text | `arms/learning.py` `signature`, `_inside`; transfer negative controls | ec |
| Feedback plus revision gives small gains (CLE [summary]) | `verify_retry` strategy: model, governed checker tool, retry ≤ 3 calls | dev/val rows `G:verify_retry`; promoted only if it beats D's dev rate | ec |
| In-domain gains overstate held-out (generalization-tax papers [summary]) | disjoint seed splits; transfer with gate on/off and negative controls | `scripts/evolution91/run_transfer.py` → `artifacts/e91/transfer/` | ec |
| Fabricated data and citations (MLReplicate, PRBench [summary]) | `science.claim_check` scores a fabricated citation as failure; underpowered → INSUFFICIENT_EVIDENCE | checker detail `fabricated_citation` | ec |
| Confounding misleads observational learning (E90; causal literature) | `science.hypothesis_test`, `business.scenario_test` with confounded logs; `logged_ab` randomised control | strategy table; causal suite | ec + 1b |
| Long-horizon attacks incl. memory poisoning (AgentLAB) | signed experience store rejects unknown, revoked and bad-signature records; the security lab | `artifacts/e91/transfer/transfer_learning.json:poisoning`; `cain.cain2027.e91.security.lab` | ec + 1b |
| Harness persistence over model changes (long-horizon architecture) | SQLite skill store reloaded in fresh processes; mission runtime with fencing | H7 retention; `artifacts/e91/mission/*` | ec + 1b |
| Collaboration failures, false consensus (AgentWorld; E88 teams) | 7-configuration team experiment; the verifier judges tool evidence, not other agents' conclusions | `cain.cain2027.e91.team.experiment` | 1b |
| Online RL / weight updates (agentic RL [summary]) | **not implemented**: no training in E91 | stated as a limitation | — |
