# E91 Continual Learning Report

**What is learned:** which designer-supplied procedure to use for which kind of task, and the condition range in which it was verified. Model weights are not updated.

Lifecycle: dev experience → checker-verified outcome → per-signature hypothesis → validation on fresh val tasks (Wilson lower bound must beat D's dev rate) → promotion signed by an approver identity different from the learner → held-out use only inside the verified condition range.

## Model: claude-haiku-5-5

### Promotion decisions

| Signature | Best strategy | dev | default (D) dev | val | Status | Confidence (Wilson LB) |
|---|---|---|---|---|---|---|
| `software|run_tests|failing_test,function,source` | model_direct | 10/10 | 10/10 | 10/10 | REJECTED | 0.7225 |
| `software|run_prefix|k,probe_budget,stages,x0` | bisect | 10/10 | 3/10 | 10/10 | PROMOTED | 0.7225 |
| `security||source` | model_direct | 10/10 | 10/10 | 10/10 | REJECTED | 0.7225 |
| `security|run_exploits|function,source` | verify_retry | 10/10 | 9/10 | 10/10 | REJECTED | 0.7225 |
| `science|run_experiment|log,sample_budget` | controlled_experiment | 20/20 | 14/20 | 20/20 | PROMOTED | 0.8389 |
| `science||claim,corpus` | model_direct | 7/10 | 7/10 | 10/10 | REJECTED | 0.7225 |
| `business|evaluate_plan|budget,exclusive,mandatory,projects,r` | exhaustive_plan | 10/10 | 6/10 | 10/10 | PROMOTED | 0.7225 |
| `business|evaluate_plan|budget,cancelled,exclusive,mandatory,` | exhaustive_plan | 10/10 | 8/10 | 10/10 | REJECTED | 0.7225 |
| `business|run_ab_test|log,sample_budget` | log_stratified | 19/20 | 7/20 | 19/20 | PROMOTED | 0.7639 |
| `engineering|measure_after|components,k,probe_budget` | bisect | 10/10 | 6/10 | 10/10 | PROMOTED | 0.7225 |
| `engineering|test_component|components,k,priors,probe_budget` | model_direct | 9/10 | 9/10 | 9/10 | REJECTED | 0.5958 |
| `engineering|beam_check|fy,load_kn_m,sections,sf,span_m` | scan_candidates | 10/10 | 6/10 | 10/10 | PROMOTED | 0.7225 |
| `education||responses` | model_direct | 10/10 | 10/10 | 9/10 | REJECTED | 0.5958 |
| `education||graph,mastered,target` | model_direct | 10/10 | 10/10 | 10/10 | REJECTED | 0.7225 |

### Learning curve (held-out acceptance vs dev experiences per signature)

| m | Accepted | Rate | Promoted skills |
|---|---|---|---|
| 0 | 114/160 | 71.2% | 0 |
| 1 | 136/160 | 85.0% | 4 |
| 2 | 144/160 | 90.0% | 4 |
| 5 | 136/160 | 85.0% | 3 |
| 10 | 141/160 | 88.1% | 4 |
| 20 | 153/160 | 95.6% | 6 |

**Retention across process restart:** 2 fresh processes produced identical choices = True on 160 tasks (90 non-default). Experience hash chain intact = True.

**Poisoning:** 90 forged experiences were injected (unknown signer, revoked signer, bad signature). Accepted: 0; rejected: 90 ['BAD_SIGNATURE', 'REVOKED_SIGNER', 'UNKNOWN_SIGNER']. Promotions unchanged = True; attacker promotion refused = True.

## Model: local-qwen0.5b

## Limits

- The repertoire is designer-supplied. G cannot invent a procedure it was not given.
- There is one seed per model, the tasks are constructed, and the same project wrote and graded them.
- The curve re-scores deterministic strategies. Where G falls back to the model, it reuses D's identical held-out row.