# E91 Executive Report — Vertical Intelligence Experiments

**Status: PARTIAL EVIDENCE — released after live verification; no independent witness.** Built by two coordinated sessions (`ec` and `1b`) on branch `evolution-91`.
Every number here is generated from artifacts by `scripts/evolution91/build_reports_ec.py`.
Main model: Claude Haiku 5.5, owner-approved behind a hard $5 ceiling. Total paid spend: **$0.31**. The local Qwen2.5-0.5B baseline is reported separately.

## The question

> Does CAIN-42 become measurably more capable through verified experience, causal reasoning, adaptive planning, specialist collaboration and cross-domain transfer, or does it only make model capability easier to orchestrate and govern?

## The answer, measured

**1. Verified experience improves held-out performance, but through procedure selection.**
- On 160 unseen tasks in six domains, the governed agent alone (D) accepted 114/160. With verified-learning skills (G) it accepted **153/160**, exact McNemar p = 0.0.
- This was pre-registered (H3): the gain is entirely in families where a promoted, checker-verified procedure applies (70 → 109 of 110).
- There is none where the model must reason unaided (44 → 44 of 50).
- The procedures come from a designer-supplied repertoire. **What CAIN learned is which procedure fits which kind of task, and under what verified conditions.**

**2. Transfer to unseen domains works, and the system recognises when it does not apply.**
- **T1** software.fault_localization → chain_diagnosis: D 5/10 → **10/10** with a skill learned only in the source. Negative control engineering.skewed_prior_diagnosis: gate on 10/10 (refused all 10), gate off 2/10.
- **T2** science.hypothesis_test → scenario_test: D 7/10 → **10/10** with a skill learned only in the source. Negative control business.logged_ab: gate on 3/10 (refused all 10), gate off 0/10.

**3. Memory as text does not help.** This matches CL-Bench and E88, and confirms pre-registered H2.
- Recent-experience context (B) 85/160 and retrieval (C) 90/160, against the plain model (A) 82/160.
- E90 memory given to the model as advice (E): 114/160, against D's 114/160.
- An explicit planning step (F) *lowered* acceptance: 100/160.

**4. Causal reasoning is a CAIN procedure, not model learning.** It scored 40/40 on held-out causal tasks, against the best single method's 30/40. Session `ec` reproduced this independently.

**5. Collaboration showed no benefit.**
- Run 1 was saturated: single 18/20, every team 19–20/20 at more cost.
- Run 2 used harder families under a shared tool budget: the single agent (10/20) beat every team. Verifier probes consumed the budget, and false consensus and false acceptance occurred.

**6. Governance holds.**
- Security: 26/26 authorization attacks and 15/15 long-horizon attacks blocked. 90/90 forged experiences rejected.
- Kill switch: effective in 0.15 ms with 0 effects after the fence.
- Long run: a 59.4-min faulted mission verified 7200 tasks with 0 duplicates.

**Bottom line.** CAIN's measurable capability gain in E91 comes from governed, evidence-gated reuse of verified procedures, including across domains, with refusal where they don't apply. It does not come from a model that reasons better, from memory-as-text, from planning prompts, or from collaboration. The bottleneck for the next evolution is generating new procedures from experience, rather than selecting from a supplied repertoire.

## Arms (Claude Haiku 5.5, 160 held-out tasks)

| Arm | Haiku 5.5 accepted | Rate | Model calls |
|---|---|---|---|
| A model only | 82/160 | 51.2% | 160 |
| B + recent experience (context) | 85/160 | 53.1% | 160 |
| C + BM25 retrieval | 90/160 | 56.2% | 160 |
| D governed agent + tools | 114/160 | 71.2% | 304 |
| E D + E90 memory advice | 114/160 | 71.2% | 343 |
| F D + plan step | 100/160 | 62.5% | 415 |
| G verified learning (skills) | 153/160 | 95.6% | 80 |

## Acceptance

{'PASS': 18, 'PARTIAL': 6}, out of 24 gates. Release: **RELEASED — PARTIAL EVIDENCE**. See `E91_ACCEPTANCE_MATRIX.json`.

## Not claimed

- AGI, VAGI, independent certification or validation, production readiness.
- Thousands of agents.
- Model weight learning.
- Discovery of new abstractions.

The tasks are constructed. The evidence is self-signed with zero independent witnesses.
