# E126.b Negative Results

## N1 — Active causal selection did not beat random

- Active MAE 0.1101 vs random 0.1336, but paired wins **25 active / 25 random / 30 ties**.
- The preregistered paired test failed. **FAIL**, despite a lower mean MAE.
- Counterfactual MAE: active 0.0768 vs passive 0.2812. **PASS** for counterfactuals.

## N2 — Novel-environment H1 not met

- 200 instances, seed 130: full **0.895** vs retrieval **0.860**, p=**0.210**.
- Full vs E125: 0.895 vs 0.820, **+7.5 points**, p=0.020; below the preregistered 10-point H2.
- The pilot seed 129 (0.88 vs 0.79) did not reproduce the margin. **PARTIAL.**

## N3 — Continual learning is not better than every memory baseline

- Full 0.6667 vs no-memory 0.5556 (+11.1, p=0.00049) — a real gain.
- Full **ties retrieval** (0.6667) and **episodic** (0.6667), p=1.0.
- The claim is only: full beats no-memory/semantic/consolidated with better retention.
- It is not a claim of universal memory superiority.

## N4 — Metacognition does not abstain

- Brier 0.0926 / ECE 0.0082, but abstention rate 0.0 at threshold 0.5.
- A well-calibrated predictor on this task is not proof of general self-knowledge.

## N5 — Adaptation retention still fails

- E126 LoRA replay target 0.5417 vs frozen 0.25.
- Control **0.375 → 0.0**. Catastrophic forgetting is not solved.
- The E126.b release gate remains FAIL for retention.

## N6 — Mission is not six hours

- E126.b mission target 600 seconds; the actual result is in the bundle.
- A process surviving 10 minutes is not long-horizon autonomy.

## N7 — External validation is blocked

- ARC-AGI, AMA-Bench, CLE, METR/HCAST/RE-Bench and official SWE-bench are unavailable here.
- External human reproduction remains BLOCKED_EXTERNAL_REVIEW.
