UNIFIED STATE · NEWS

Carrier’s relationship editor passes its next test

A new small learned specialist improves the connections in Chromatic Habitat replies. On 1,200 fresh numerical scenes, typed-link F1 reached 0.9912, compared with 0.9018 for copying and 0.5529 for the earlier model on those same scenes. Three independent training seeds produced closely similar results.

The new component makes selective decisions about existing links instead of rebuilding the whole list. It adds 11,055 trained parameters to the frozen earlier field model. Other output channels remain inherited.

The result has an important boundary: complete link sets were exact in 76.7% of cases, and F1 on genuinely new birth connections was only 0.5255. Success at preserving and retyping existing relationships is not success at every form of invention.

An exploratory test near the teacher’s exact distance thresholds exposed another failure: link F1 was 0.600, no better than copying, and only one third of links that should persist were retained. The improved model is not universally dependable.

The public instrument includes editable compositions, raw model comparisons, optional visible protection, art trajectories, reproducible weights and an independent metric audit. Its results concern the stated authored exercises, not semantic understanding or subjective experience.

Open Chromatic Steward · Read the method, limitations and evidence