openmoji-g1-slot-bound-v5-f502df1-d8af55ea-9b9b1699

Written result

The encoder summed six coordinate lookups — all from one shared embedding table — into a single vector per segment slot. A sum is commutative, so a segment was represented by an unordered bag of its coordinate values. In float64, a program and its coordinate-swapped variant produce byte-identical logits: maximum absolute difference exactly 0.0 across 414 swapped segments in six held-out icons, with identical argmax predictions. The model provably could not tell which coordinate held which value, while still having to predict all six separately.

This run binds each value to its slot by elementwise multiplication with a learned per-slot vector. Everything else is the v2 configuration exactly: same split, seed, fixed corruption 0.35, batch size, learning rate, step cap and selection policy. Noise conditioning is off, since v4 falsified it. The change costs 768 parameters, +0.13%.

Result

Early stopping ended the run at step 1,980; the selected checkpoint is step 1,500 — the latest optimum of any run so far. A second complete invocation returned an identical JSON result.

768 parameters beat a 3.53x capacity increase on every measure, which is the clearest confirmation that the defect was real and structural rather than a matter of resources. Retained-token accuracy rises 31%, aggregate accuracy 31%, changed-token recovery 37%, and held-out loss improves for the first time since v2.

Predeclared criteria

Standing Gate G bar, unmet since v1: retained-token accuracy >= 0.90, observed 0.5408.

Overall: partially falsified.

The gain is real; the mechanism is not fixed

Every recovery figure improves, and at corruption 0.20 the model is now essentially break-even where v2 lost ground, helping on 20 of 32 icons rather than 14. But the correlation between its predictions at 0.05 and at 0.35 is unchanged — 0.8621 against v4's 0.8619 — and mean x_hat_0 error still moves only from 0.124 to 0.143 across a sweep where the input moves from 0.051 to 0.180.

So the model reads its input better than it did, and uses it barely more. Fixing the encoder raised the ceiling of what it produces without changing what it fundamentally does: emit a prior lightly adjusted by its input.

What remains

The predeclared falsification meaning holds, and it is now the only candidate left standing after data volume, capacity, the corruption regime, noise-level information and encoder slot-blindness have each been eliminated by a predeclared comparison:

The prediction objective itself. The model predicts every field through an independent single-shot softmax over a 289- or 417-way vocabulary. There is no cheap way for it to express "leave this one alone" — copying a field requires reconstructing its exact token from scratch through the transformer, and at 65% of fields uncorrupted that is most of the work it is being asked to do. An edit-mask or residual formulation, where copying is the default and the model predicts only what to change, is the change that addresses this directly.

Slot binding should be kept regardless: it is a strict improvement for 768 parameters and it removes a defect that would have confounded every later result.

Scope

Fixed-topology and geometry-only, single seed, one corruption draw per icon at each sweep level. No run in this sequence has yet produced a recognizable icon, and this one does not either. It is a better-conditioned model, not a working one.