openmoji-g1-corruption-sweep-71a080a-4levels-9b9b1699

Written result

v2 eliminated data volume as the Gate G constraint and v3 eliminated model capacity, both on matched predeclared comparisons. The corruption schedule was the only candidate left, and the render probe had suggested why: at probability 0.35 the input x_t is already visually destroyed, so perhaps no model can recover information that is gone.

This run tests that before spending any training on it. One fixed checkpoint — v2's loss-selected step 840 — evaluated and rendered at four corruption levels, with 0.35 as the control because that is what it trained at. Identical icons, sizes and corruption seeds at every level; only the Bernoulli rate differs. No training.

The statistic is the per-icon recovery fraction, (MAE(x_t) - MAE(x_hat_0)) / MAE(x_t) — the share of the render gap the model closes. Absolute error is not comparable across levels, because x_t itself moves closer to x_0 as corruption falls.

Result

At 18 px the picture is the same: -4.07, -1.12, -0.12, +0.19.

Predeclared criteria

The regime hypothesis is falsified and the competing damage hypothesis — registered in the same record before launch — is confirmed. At probability 0.05 the model makes its input 2.8x worse, and it helps on zero of thirty-two icons.

Why: the output barely depends on the input

The decisive number is not in the table above. It is that x_hat_0 error hardly moves at all while x_t error moves by a factor of 3.5:

Per icon, the prediction at p=0.05 correlates with the prediction at p=0.35 at Pearson r = 0.91, with a mean absolute difference of 0.019 against a mean error of 0.14. The model emits nearly the same reconstruction for a given icon whether 5% or 35% of its geometry has been replaced.

It is not denoising. It has learned a group- and subgroup-conditioned prior over icon geometry and emits approximately that, largely ignoring x_t.

The architecture makes this easy to believe. GeometryDenoiser takes the corrupted tokens plus group and subgroup embeddings and nothing else: there is no noise-level or timestep input and no mask marking which fields were replaced. Trained at a single fixed probability of 0.35, it never needed to behave differently at any other level, and nothing in its input tells it how much to trust what it is given.

This corrects how the earlier Gate G numbers should be read

The measurements stand; their interpretation does not.

• The +18% recovery at p=0.35 in the first render probe is not the model recovering geometry. It is a fixed-quality output happening to beat a badly corrupted baseline. The same output scored against a lightly corrupted baseline loses badly.
• Held-out retained-token accuracy of 0.4128 was never a secondary weakness. It was this finding, visible in token space the whole time: the model overwrites correct fields because it is not reading them. The standing 0.90 bar has been measuring the real problem since v1.
• The held-out loss floor near 7.35 that v2 and a 3.53x larger v3 both converge to is consistent with both models learning the same prior. Capacity would not help, and measurably did not.
• v1's 18x-above-untrained result and v2's 25% loss reduction remain real learning. What was learned is a prior over icons, not a conditional denoiser.

What this redirects the project toward

Not a corruption schedule. The falsification meaning was registered before launch and it holds: the binding problem is that the model cannot leave correct fields alone, so the next experiment is an objective or architecture that lets it.

The concrete, cheap candidates, in the order their cost suggests:

1. Tell the model the noise level. A timestep or corruption-level embedding, and training across a range of levels rather than one fixed point. Without it the model cannot modulate how much to trust its input, which is exactly the failure here.
2. Make copying easy. Predict an edit mask, or a residual against x_t, so that "leave this field alone" is the default rather than something the model must reconstruct token by token.
3. Condition on which fields were corrupted, as a diagnostic upper bound. That is not a realistic sampling-time signal, but it cleanly separates "cannot identify the corrupted fields" from "cannot predict their values", and that is worth knowing before designing around either.

The first is the smallest change that addresses the mechanism observed here, and it is the one this project should run next.

Scope

One checkpoint, one seed, 32 held-out icons per level, one corruption draw per icon. The model trained only at 0.35, so evaluating it at 0.05 is off-distribution by construction — that is the point of the sweep, but it means a poor result at low corruption does not by itself condemn a model that was trained across levels. Establishing that is exactly what candidate 1 above would do.