# E91 Next Evolution Recommendation — E92

E91 established a governed evaluation harness and a real, measured gain from **verified procedure selection**, including cross-domain transfer with refusal. It also produced clear negative results. E92 should attack the bottlenecks those results exposed, not add breadth.

## 1. The central bottleneck

E91 G improved from 114/160 to 153/160 by selecting from a **designer-supplied procedure repertoire**. It did not invent a procedure. The next capability is:

> Given a new task family and verified outcomes, generate a candidate procedure, test it, and promote it only when an independent checker and held-out validation support it.

Success criterion: on held-out families with **no supplied procedure**, G2 must beat D and the best fixed strategy at p < 0.05, and the generated procedure must be inspectable and reproducible.

## 2. E92 experiment set

| # | Experiment | Minimum evidence |
|---|---|---|
| E92-1 | Procedure synthesis from verified examples (no supplied repertoire) | Held-out generation families; G2 > D; negative controls; generated procedure shown |
| E92-2 | Adaptive reasoning completed | Full 160-task H run; static vs adaptive with equal compute; cost and latency reported |
| E92-3 | Collaboration under a realistic budget | At least one family where a specialist team beats a single agent; false acceptance and consensus reported; if none, publish the negative result |
| E92-4 | Real repositories and isolated security targets | At least 20 software tasks on unseen repositories and 20 authorized isolated targets; independent reference tests |
| E92-5 | Long-horizon 6 h and 24 h | Only claim elapsed time actually completed; 72 h remains future unless a real run exists |
| E92-6 | Independent verification | One external verifier re-runs the bundle and at least one experiment; publish the verifier's identity and result |
| E92-7 | ClawX API integration | Missions written into the real ClawX data model, tenant-isolated and offline-exportable, not just a static page |
| E92-8 | Local model baseline completed | The same 160 tasks on a local model, or a disclosed reduced set; no mixing of local and paid rows |

## 3. Method requirements

- Keep the deterministic checker as the only acceptance authority.
- Keep held-out seed ranges disjoint from dev/val.
- Pre-register every comparison and the stopping rule.
- Report duplicate rows, incomplete arms, and excluded tasks.
- Keep paid spend behind a hard ceiling; if no credit remains, mark the paid comparison BLOCKED, not passed.
- No production deployment or customer claim without authorization and evidence.

## 4. What not to do

- Do not claim model learning: E91 trained no weights.
- Do not claim collaboration benefit from a saturated run.
- Do not call a 59.4-minute run a one-hour mission.
- Do not call a self-signed bundle independently verified.
- Do not present a designer-supplied procedure as discovered intelligence.
- Do not add six more verticals before the procedure-generation bottleneck is measured.

## 5. Recommended priority

1. **E92-1 procedure synthesis** — the actual intelligence bottleneck.
2. **E92-2 adaptive reasoning completion** — finish the arm that E91 left at six rows.
3. **E92-4 real repositories** — the fastest way to make the result matter outside the simulator.
4. **E92-6 independent verification** — the trust bottleneck.
5. Everything else after those four.
