CHROMATIC SEED: RELEASE DISCLOSURES

The training and final predictive split were frozen before fitting. The matched-seed reply counterfactual was specified separately after predictive metrics were known, before processing replies. It is exploratory input-sensitivity analysis, not a preregistered semantic result.

The small network receives categorical indices allocated to exact RGB values. It does not receive natural-language definitions or numeric RGB channel geometry. Carrier's existing graph carries author-supplied structural information. This is not an unconditioned blank slate.

Interpreter tasks excluded the training topology file, other encounters, earlier artworks and predictive metrics by instruction. The selected definition records DID retain existing relation/provenance metadata. Thus the interpreters had some inherited relational information. They were fresh pretrained agent contexts, not OS-isolated or independently authenticated model builds. No claim of graph-blinded interpretation is made.

Evaluation file keys 'target_unseen_in_training' mean absent from scored training TARGETS. Of those 199 held-out queries, in 194 the target identity had occurred as a scored training INPUT. The 199 queries concern 192 distinct target identities; the 194 input-seen queries concern 188 distinct identities. They are not all wholly unseen colours.

Training completed1200 updates; the selected checkpoint is update200. One model run is insufficient for population claims. Colour use, reply effects and first readings are exploratory descriptive records, not validated artistic-quality scores.

The small model processes only non-null reply colours in stored cell-array order. Nulls, geometry, depth, emphasis, typed links, regions and explanation are omitted from its input, with omission counts preserved. Model weights are frozen during exchanges. Changed output is not proof of understanding; unchanged samples are retained.

Prior public authorship/provenance within the frozen Carrier snapshot is retained. No private conversation is included.

Initial runtime setup needed a workspace temporary directory and permitted package downloads. Failed temp/network attempts remain in setup. The first renderer-copy command used the wrong relative path; it stopped generation before encoding. That helper error is preserved. The three untrained PNG footers initially used the common learned-field caption; corrected versions explicitly say untrained. Original labelled images remain in setup/caption-first-renders. Raw colour fields and model parameters did not change.

All local numerical/static/inert checks are reported as such. No physical phone, real browser rendering or live WordPress deployment is claimed. The kit is prepared for upload; it was not deployed by this task.