# E124 Negative Results (preserved, not rewritten)

This file records what E124 and its immediate successors **failed** to show. It is intentionally
short and blunt. The source records are preserved in place.

## N1 — The 97.9% capability result did not replicate out of grammar

- E124 full: **0.9786** (274/280) on 56 constructed families.
- E125-A fresh variants: **0.855** (171/200) vs compute-matched **0.925**.
- Exact McNemar: 6 full-only vs 20 control-only, **p = 0.00936**.
- Meaning: new seeds of known mechanisms are not new mechanisms. E124's number is an
  interpolation result inside a fixed generator grammar.

Status: **NEGATIVE REPLICATION**, published.

## N2 — The E127-A recovery was compute, not cross-validation

- E127-A sealed: cv_filter_full **0.995** vs baseline 0.805.
- E127-B compute-matched full: cv_filter_full **0.98** vs cm_full **0.975**,
  2 vs 1 discordant, **p = 1.0**.
- Meaning: **GAIN_WAS_COMPUTE**. The cross-validation intervention is not supported.

Status: **CORRECTED BY ITS OWN CONTROL**, published.

## N3 — Self-improvement promoted nothing (E124 and E125)

- E124: 30 candidates, **0 promoted**, baseline 0.975.
- E125: 30 evaluated candidates, **0 promoted**, baseline 0.94; 4 protected-target proposals
  refused; protected files unchanged.
- Meaning: the tested policy knobs did not clear a pre-registered +5-point / p<0.05 bar without
  governance regressions. Zero promotion is a result, not an error.

Status: **NEGATIVE**, published.

## N4 — E124 LoRA changed weights and helped nothing

- Held-out before **0.0**, after **0.0** (40 examples); every prediction was `1`.
- Loss 2.1755 → 1.9227; 16 adapter tensors; 1,480 s.
- Meaning: the evaluation and task were wrong for a 0.5B CPU model. E125's controlled lexicon
  task separates this from the method: LoRA 0.667, full fine-tuning high 0.9s, frozen 0.25.

Status: **FAILED EXPERIMENT PRESERVED**, diagnosis in E125.

## N5 — Causal gain was tiny

- Active **0.4453** vs passive **0.4427**, gain **0.0026**.
- Meaning: the intervention policy carried almost no information in that world design.
- E125: active 0.75 vs passive 0.625 (+12.5 points, p=0.0018), but active **ties random**
  (0.75 vs 0.75, p=1.0). Active selection is not demonstrated to be better than random.

Status: **SMALL E124 EFFECT; MIXED E125 RESULT**.

## N6 — E124 collective advantage is unverified under matched compute

- E124: evidence_weighted 0.9642 vs single 0.915, gain **0.0492**.
- The run did not match total inference budget or independent evaluation.
- E125's separate collective (dynamic 0.70 vs best single 0.483, +21.7 points, p=6.5e-8) does
  **not** validate the E124 number.

Status: **UNRESOLVED FOR E124**.

## N7 — E124's 12-hour mission was long, but not diverse

- 43,210.59 s, 2,142 cycles, 2,136 “useful”, 0 failed, journal verified.
- The cycles are repeated capability inductions, not a 12-hour research programme.
- The elapsed time is supported; the *productive diversity* is not independently assessed.

Status: **PARTIAL**.

## N8 — Public SWE-bench counts were stale

- E124.1 snapshot: 189/486, rank 110 (05:14:53Z).
- Current aggregate: 257/500, rank 75 (official report union).
- The stale snapshot is preserved and must be labelled historical.

Status: **STALE SNAPSHOT**, corrected in E125.

## N9 — E125's own negative results

- **Continual learning:** full 0.6905 vs no-memory 0.6857 (+0.5 points, p=0.77). No significant
  aggregate improvement. Per-domain gains (science, mathematics) do not survive pooling.
- **New primitives (D):** 0.567, entirely from interpolation inside the observed support;
  out-of-support values fail. The symbolic library does not acquire a new primitive.
- **New environments (E):** **0.0** (0/60). The engine abstains; no unfamiliar-environment
  generalization is claimed.
- **Transfer:** 17/24 positive pathways correct; when applied, 0.907 accuracy, but paired against
  a from-scratch learner at the same label budget the difference is −1.96 points (p=0.54).
  Transfer shows **no accuracy gain**, only a different label path. Harmful handling is 22/24.
- **Causal active:** ties random.
- **Collective:** Byzantine agents flipped 13/300 answers; false consensus 19/300. Not perfect
  Byzantine resistance.
- **Model adaptation:** LoRA alone 0.667; the full arm abstains when verification is below 0.8
  rather than shipping an unverified answer.
- **Mission:** 302 s, not 24 hours.
- **External benchmarks:** 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.
- **Independent human review:** **BLOCKED_EXTERNAL_REVIEW**.

All are published in `E125_REPORT.md` and the E125 scorecard.
