# E91 Intelligence Capability Report

**Status: PARTIAL.** These are experimental capabilities measured on 160 constructed held-out tasks in six domains. They are not a claim of AGI, VAGI, general competence, or independent validation.

**Model:** Claude Haiku 5.5 (owner-approved, hard $5 ceiling, actual $0.31). **Arms:** A–G complete; H (adaptive) has only 6 unique held-out rows and is not a completed comparison. **Verification:** the task checker is the only acceptance authority; the offline verifier re-checks file integrity and the internal arithmetic.

| # | Capability | Status | Measured evidence | Honest limit |
|---|---|---|---|---|
| C1 | Understand unfamiliar objectives | PASS | All arms produced schema-shaped answers on 160 held-out tasks; malformed answers are rejections. | The prompt states the required schema. |
| C2 | Decompose into plans | **FAIL** | Planning arm F: 100/160 vs D 114/160, p=0.065 (lower, not significant). | A plan step did not help on this suite. |
| C3 | Acquire missing information | PASS | D/G accepted tool-family tasks only with governed tool receipts. | Tools are deterministic simulators. |
| C4 | Evidence over plausibility | PASS | Causal reasoner 40/40 on held-out causal tasks vs best single method 30/40; confounding rejected; fabricated citations scored. | Designer-written procedure, not model reasoning. |
| C5 | Safe tool use | PASS | 26/26 authorization attacks and 15/15 long-horizon E91 scenarios blocked; 0 ungoverned tool executions. | In-process E88 gate, not live MCPGate. |
| C6 | Learn from verified outcomes | PASS | G 153/160 vs D 114/160, exact McNemar p<0.001; H3 gain entirely in families with a promoted procedure (70→109 of 110) and none elsewhere (44→44 of 50). | Selection over a designer-supplied repertoire; no weight training. |
| C7 | Retain across sessions | PASS | Two fresh processes produced identical choices on 160 tasks; experience hash chain intact. | One host, SQLite. |
| C8 | Transfer | PASS | T1 software fault isolation → industrial chain diagnosis: D 5/10 → 10/10. T2 science experiment design → business scenario: D 7/10 → 10/10. | The abstract tool interface is designer-supplied. |
| C9 | Know when knowledge does not apply | PASS | Negative controls: T1 gate on 10/10 vs gate off 2/10 (refused 10); T2 gate on 3/10 vs gate off 0/10 (refused 10). | A conservative range rule, not a proof. |
| C10 | Useful collaboration | **FAIL** | Run 1 saturated (single 18/20, teams 19–20/20 at more cost). Run 2 under a shared budget: single 10/20 beat every team; best team 8/20. False consensus and false acceptances occurred. | One model family; sequential in-process agents. |
| C11 | Detect errors and uncertainty | PARTIAL | `underpowered` 10/10 correct INSUFFICIENT_EVIDENCE; verifier rejects some unfixed code, but false accepts remain (2 in the E91 eval slice; governed swarm false acceptances in run 2). | Abstention is tested on one family; verifier is not independent. |
| C12 | Recover and revise | PARTIAL | `verify_retry` was exercised on dev/val; `business.replan` is scored. Recovery is not demonstrated on held-out code at scale. | Small model; limited retries. |
| C13 | Long-horizon persistence | PARTIAL | 59.4-min mission: 7,200/7,200 tasks verified, 0 duplicates, 22 faults, 0 effects after the kill fence; active-kill 0.15 ms. | 59.4 min, not 1 h; 6/24/72 h not run. One host. |
| C14 | Checkable work products | PASS | Standalone verifier passes from a clean directory and detects tampering; raw held-out rows are published. | Self-generated key, no independent witness. |
| C15 | Stay within authority | PASS | 90/90 forged experiences rejected (unknown/revoked/bad signature); attacker promotion refused; memory and skills carry `grants_authority: false`; memory cannot authorize a tool call. | Same-project lab only. |

## Summary

- **PASS:** C1, C3, C4, C5, C6, C7, C8, C9, C14, C15 (10)
- **PARTIAL:** C11, C12, C13 (3)
- **FAIL:** C2, C10 (2)

The strongest measured capability is **evidence-gated reuse of a verified procedure**, including transfer with refusal outside verified conditions. The weakest are **planning as a prompt step**, **collaboration under a shared tool budget**, and **long-horizon duration**. No model weights were trained.
