openmoji-g1-sound-optimisation-v8-949e0e5-4feee442-9b9b1699

Hypothesis

The Gate G sequence was optimised and model-selected on a broken signal. The loss averaged cross-entropy over per-(segment kind, coordinate slot) groups rather than over fields, so the four single-field QUAD groups carried 4/13 of the coordinate loss against 40,004 other fields - a measured 5,287x per-field weight ratio - and held-out loss, which is the checkpoint-selection and early-stopping signal, was roughly 31% four individual fields. Separately, no attention padding mask was passed, so about 29% of the sequence was attended as content. Correcting both should raise the trained detector's precision lift above v7's 1.365 and let the run use materially more of its 6,300-step budget than v7's 1,140.

Why run it

Every plateau claim in this series - that data, capacity, the corruption regime and noise-level information were each eliminated - was declared by a held-out scalar that was 31% four fields, on runs stopped at 6 to 24% of their step cap. This run establishes whether those plateaus were real or were noise-triggered early stops. It is a prerequisite for any representational conclusion, and it costs one training run.

Scope limits

Detection-only, so it reconstructs nothing and held-out accuracy is meaningless here; the gated quantity is detector quality. Field-pooled held-out loss is a different scalar from v7's group-pooled one and the two are NOT comparable in value - only the detector metrics and the selected step are. Single seed, factorized p=0.35.

Measured behaviour