Soft cross-rendering convergence vs. graph-verified carrier-independence, and the value-erased-schema control boundary

Summary

This study tests a carrier-independence premise for a State-over-Token modeling approach: the hypothesis that a learned next-state predictor projects a single situation, rendered in structurally distinct carriers (free text, a visual canvas, a structured trace, a symbolic schema, and a surface-corrupted variant), into one shared semantic state. Two increments are reported: (a) transplanting a previously developed multi-encoder convergence mechanism from an original kinship micro-world onto a structurally distinct logistics-provenance world, and (b) adding a schema-rendering encoder and a noisy-rendering falsifier on top of a validated three-rendering baseline.

A cold external review prompted a scope correction. The convergence demonstrated here is the soft form — embedding/retrieval-band agreement among controlled symbolic renderings. The hard form (learned recovery of a shared graph across renderings) sits at the shuffled-control floor (isomorphism rate near zero; a matched pairing is indistinguishable from a random one). Strong carrier-independence is therefore not verified, and natural multimodal grounding is untested — every result is a hand-authored symbolic cell.

Increment 1 — cross-world transplant (single-seed, inconclusive)

A single-seed full training run was executed on the logistics-provenance cell. The strict three-test gate did not close under the single-seed protocol, so this is treated as a partial-transfer finding with directional information rather than a falsifying negative. Three structural observations:

1. Text-encoder mechanism transfers. Text accuracy on the one-hop relational query reached roughly 0.9, versus roughly 0.2 for the comparable query in the predecessor world — an approximately fourfold improvement. Because both tasks are one-hop relation queries over small entity sets, the gain reflects the cross-rendering reasoner mechanism working when the entity-ID pathway is healthy, not an inherently easier task.

2. Entity-ID descent transfers. The entity-ID loss descended to a low value, whereas in the predecessor world the same loss ascended into a high range — the load-bearing prior-collapse failure of the earlier chapter. With the same architecture and same loss weights but a denser, lower-vocabulary world, the collapse disappears. The conclusion is that the prior-collapse failure was not mechanism-fundamental; it was a property of a high-vocabulary, Zipfian-frequency setup. The denser cell removes the collapse. This is catalogued as a vocabulary-density operating boundary.

3. The visual encoder becomes the dominant bottleneck. Visual accuracy lagged text accuracy by roughly a factor of two — a gap absent in the predecessor, where text and visual both sat near the noise floor. Candidate causes (more entity kinds, denser canvas layout, purely positional per-entity glyphs) are design targets for follow-on work, not architectural failures.

Argmax-floor persists

Rendering-isomorphism stayed near zero even though the dense logits converged. The transplant cured the prior collapse, but the combinatorial decode (slot-to-ID assignment, edge usage, source/destination slot assignment) still does not align across encoders. This re-confirms an earlier small-cell "argmax-floor" finding as a deeper property of the rigid slot-based discrete state schema rather than a cell-specific artifact, and motivates a latent-schema research direction.

Increment 2 — schema + noisy carriers (partial-positive closure)

Two new carriers were added on top of the three-rendering baseline: a schema view (a symbolic, value-erased structural skeleton of entity slots and relation slots, routed through the shared text-encoder backbone to avoid an encoder-capacity confound) and a noisy view (surface corruption: typos, word-drop, visual occlusion; calibration details abstracted) added at observe-only weight as a falsifier.

A wiring-integrity trap (governance note)

An earlier version of the trainer extended the loss function to accept the new schema/noisy arguments but never overrode the training step. The inherited step called the loss positionally, so the new arguments stayed unset and the schema convergence term collapsed to a no-gradient zero — schema-augmented training was a complete no-op, and that version's closure numbers were invalidated (its checkpoints are functionally plain retrains of the prior increment). The fix overrides both training and validation steps to render and encode the schema view every batch, and tokenizes schema strings before the shared encoder. A short smoke test confirmed nonzero schema-text convergence, nonzero-but-zero-gradient noisy terms at observe-only weight, an encoder gradient-norm that rose with the convergence weight, and a total-loss delta exactly equal to the sum of the schema convergence terms. This integration-hygiene trap is recorded in the decision log.

Multi-config, multi-seed results (2 configs x 3 seeds)

Baseline strengthened, not just preserved. All three-rendering convergence deltas versus the prior baseline were positive (small but consistent across seeds and configs). Schema augmentation does not destroy the baseline; it slightly strengthens it.

Reframing the noise falsifier as semantic invariance

The original falsifier expected noisy-to-clean convergence to drop substantially when noise destroys gold semantics. Empirically the sign reversed — noisy renderings were more convergent than clean ones. The ruling is that the original framing was wrong, not the result: the noise calibration is surface-destructive but semantic-preserving, scrambling surface tokens while leaving the underlying gold relations recoverable by the shared semantic state. A healthy encoder should recover, and that recovery is exactly the observed negative drop. The hypothesis is split into (a) semantic-preserving noise invariance — PASS, positive evidence of a rendering-invariant state robust to surface corruption — and (b) a genuine semantic-breaking falsifier (entity swap / relation inversion / value contradiction), deferred to future work.

Value-erased-schema control boundary

The matched-vs-mismatched schema negative control produced a near-zero margin and did not fire. This is structural, not a training failure: the schema renderer strips concrete values, emitting only the structural skeleton. A "matched" schema and a "mismatched" schema over the same situation are therefore structurally identical — they differ only in the values the renderer deletes, so there is no information left for a value-swap to perturb. The control cannot fire by construction.

Boundary finding. A schema-rendering function that strips concrete values from a situation — emitting only a structural skeleton of entity slots and relation slots without value bindings — cannot, by construction, support a value-mismatch negative control; matched and mismatched schemas are structurally identical, and the distinguishing information is precisely what the rendering deletes. Margins for such schemas are bounded near zero and indicate a schema-expressiveness limit, not a training defect. The fix — a value-sensitive schema that retains enough value channel to make the control fireable — is deferred.

Verdict

Partial-positive closure. Three strict hypotheses pass across all six runs (schema convergence, convergence floor, and baseline preservation, with the baseline slightly strengthened). Two hypotheses are reframed into principled findings rather than failures: the noise falsifier becomes positive semantic-invariance evidence with a named successor, and the matched/mismatched control becomes a schema-expressiveness boundary. Not a clean success, but no destruction, and two new boundary findings disclosed.

Cross-track relevance

The argmax-floor and visual-bottleneck findings strengthen the case for a latent-schema research direction: a rigid discrete decode never aligns even with a healthy dense representation. Future work: a semantic-breaking falsifier, and a value-sensitive schema.

Evidence level: the cross-world transplant is experimental and inconclusive (single-seed); the schema/noisy increment is reproduced across two configurations and three seeds for the three passing hypotheses.