# E126 Negative Results

Failures are kept. E126 did not achieve AGI and did not meet every preregistered target.

## N1 — Persistent learning failed again

- Full **69.6%** vs no-memory **83.3%** on 102 scored episodes.
- The pooled-rule compatibility test was too weak; memory hurt several domains.
- The preregistered target was a ≥10-point gain. The result is a **13.7-point loss**.
- Saving a skill object is not learning. This gate is **FAIL**.

## N2 — Active causal selection did not beat random

- Intervention MAE: active **0.0127**, random **0.0002**, passive **0.0624**.
- Active beat passive, but random was better on the paired worlds.
- Counterfactual predictions were implemented and scored (active MAE **0.0191**, passive
  **0.1120**), but random was better there too (**0.0004** because the design gives balanced
  interventions cheaply).
- The preregistered active-vs-random target is **FAIL**.

## N3 — The novel-environment result missed its primary retrieval threshold

- Full **87.0%** vs retrieval **78.0%**, exact McNemar **p=0.0636**.
- The preregistered H1 required p<0.05. It is **not met**.
- Full vs the E125 engine (61.0%) is **p=3e-8**, and full vs random (72.0%) is **p=0.00027**.
- The honest claim is: large improvement over E125 and random, but not statistically
  distinguishable from the retrieval control at the preregistered threshold.

## N4 — Collective verification eliminated flips but did not improve accuracy

- Verified dynamic **69.0%** vs unverified dynamic **68.4%** (+0.6 points, p=0.25).
- Byzantine flips: **0/500** (down from 13/300 in E125).
- Verification banned 22 agents (honest and Byzantine) and reduced the team's raw accuracy.
- Zero fabricated-evidence flips is a real control result; the accuracy gain is **not
  significant**.

## N5 — Mission is not six hours

- Actual duration is reported from the journal. It is not 6 h unless the journal says 6 h.
- The E126 run is a ~15-minute restart-recoverable mission, not sustained six-hour autonomy.

## N6 — Adaptation and forgetting

- The replay adaptation is reported separately in `adaptation_results.json`.
- If the control task still drops beyond the preregistered 10-point limit, the target gains do
  **not** make the update a success. The gate is PARTIAL/FAIL, not PASS.

## N7 — External evaluation is blocked

- ARC-AGI-2/3, AMA-Bench, CLE, METR/HCAST/RE-Bench, and official SWE-bench are
  **BLOCKED_UNAVAILABLE** in this build.
- No internal benchmark is relabelled external.
- External human reproduction is **BLOCKED_EXTERNAL_REVIEW**.

## N8 — Constructed and same-team

- The environment generator, agent, controls, oracle and verifier were written by the same
  three-person team.
- The SHA-256 commitment proves the answer key did not change; it does not prove independent
  witnessing.
- Seed 127 was invalidated by a post-commitment code change and is not used as confirmatory.
