openmoji-g1-retest-eliminations-b949fc2-3arms-9b9b1699

Written result

Four conclusions in this gate — that data volume matters, that model capacity does not, that the corruption regime is not the constraint, and that noise-level conditioning changes nothing — were each declared on a held-out scalar that was roughly 31% four individual fields, on runs stopped at 6–24% of their step cap, under a corruption process whose replacements a zero-parameter marginal detector identified at 2.84 lift, with coordinates encoded as unordered categories, and graded on exact-token accuracy which v15 and v16 showed points the wrong way.

All of that is now fixed, and none of the four had been retested. Three of them are training factors and are retested here off the v16 base, one factor per arm, judged on paired per-icon render recovery — the primary metric v15 and v16 established.

Result

Applying the criteria exactly as predeclared, against the base interval's half-width of 0.1103:

All three original conclusions survive. The eliminations were right even though they were measured badly, and the four are admissible again.

The data-volume arm is the sharpest of the three: training on 256 icons instead of 2,425 cuts render recovery from +0.2511 to +0.0766, a 70% drop, with its interval barely clearing zero. Under the broken measurement v2 found the same thing in exact-token terms; under the corrected one it is larger and clearer.

Capacity is the one I expected to move and it did not. v3 concluded that 3.53x parameters changes nothing, and I had recorded that as the least trustworthy of the four because 768 parameters of slot binding later beat it outright. At 3.69x on a sound setup it is still within noise of the base — and the model that does best here remains the smallest one built this session. The slot-binding result and the capacity result were never in tension: one was a representational fix, the other was raw size.

A failed arm, kept

The first noise-conditioning arm set corruption_probability to 0.05 as the low end of its training range, which also moved the evaluation level, because evaluation_corruption_probability was unset and defaults to it. Its identity baseline came out at 0.9501 against the other arms' 0.6515 — it measured a different task entirely. It is recorded in the registry as failed with that reason rather than quietly rerun, and the corrected arm pins the evaluation level to 0.35.

A second recording note: on first reading these numbers I labelled the data-volume arm as overturning v2, by applying a symmetric "differs by more than the half-width" rule to a criterion that was predeclared directionally — below the base by more than the half-width restores v2. The criterion as written is the one applied above.

Scope

Three arms, not four: the corruption-regime sweep is an evaluation-time study over a fixed checkpoint rather than a training factor, so it belongs in separate work. Fixed topology and geometry only, single seed, 12 icons for renders. No render in this project is a recognizable icon.