This is not a claim that E91 is complete. It is a local, deterministic evaluation of six vertical task families and ten decision arms. The suite is saturated by a hand-written expert policy; the report says so.
| Arm | Accepted | n | Rate |
|---|---|---|---|
| A | 26 | 120 | 0.217 |
| B | 26 | 120 | 0.217 |
| C | 22 | 120 | 0.183 |
| D | 96 | 120 | 0.800 |
| E | 120 | 120 | 1.000 |
| F | 120 | 120 | 1.000 |
| G | 120 | 120 | 1.000 |
| G0 | 26 | 120 | 0.217 |
| H | 120 | 120 | 1.000 |
| I | 120 | 120 | 1.000 |
| J | 120 | 120 | 1.000 |
A verified-experience loop raises accepted tasks from 26/120 (21.7%) to 120/120. But E and F also reach 120/120, so the gain is selecting a competent hand-written policy, not model learning. Collaboration (I) equals F and adds cost; two self-verified outputs were rejected by the checker.
Verify: python3 scripts/evolution91/eval/verify_e91_eval_offline.py .