Unified State / Research and tools
Chromatic Seed: a small model trained on Carrier colours
A learned colour generator, six interpretive encounters and a local browser instrument extend the Chromatic Habitat experiment.
Unified State has prepared Chromatic Seed — A learned habitat: a small recurrent model initialized from random weights and trained on numeric Carrier colour identities arranged through the existing concept graph.
The model has 625,745 parameters and a vocabulary of 6,257 checked colours. No concept definitions, essays or pretrained language-model weights entered its training signals. Carrier’s authored connections supplied the structure; “colour” here means an exact categorical identity rather than a measured sensation or a rule about similar hues.
The release includes nine learned fields, three untrained controls, six pretrained interpreters’ readings and replies, and the small model’s six returns. Same-seed counterfactual continuations are preserved to make the effect of changed colour inputs inspectable. There is no hidden English original to recover.
The browser tool runs the selected model locally. Visitors can choose colours, generate a 24-cell field, inspect its numerical state, reveal definitions, and save the resulting field and run record.
The technical result is mixed. Training modestly improved held-out average predictive loss, but the model’s 2.40% top-ten rate was well below the simple two-step graph baseline’s 20.66%. The release preserves this limitation alongside the art.
Replies enter the small model through their non-null colour sequence. Spatial relationships and explanatory prose remain outside that input channel. No claim of consciousness, semantic understanding or general superiority follows from the experiment.
All scheduled outputs, weights, code, frozen inputs, interpretations, figures and checks are available. Numerical and inert interface checks were run; physical browsers, phones and the live WordPress installation were not verified in this build.