12.5.3 Factored subskills

12.5.3 Factored subskills

The diagnostic repair represented active probe selection, noisy hypothesis admission, and typed extension rendering as three separately exercised skills under registered validators. Only the probe-selection policy required an edit: it learned to choose the largest registered proposal-separation score. The admission and rendering policies were already reliable once their inputs and output types were isolated.

The reported result is the separately registered v2 diagnostic. Version 1 stopped at a string-versus-number schema failure; v2 clarified the numeric JSON contract without relaxing the validator, and the records are not pooled as repeated trials.

On fifteen fresh admission worlds, the initial subskills achieved \(10/15\) end-to-end accuracy and made \(3/15\) false extensions. The frozen factored pipeline achieved \(15/15\) end-to-end accuracy and no false extensions. Probe-label imitation itself was only \(10/15\), yet all selected probes were decision-sufficient. This distinction matters: the correct target is an equivalence class of sufficient, cost-aware experiments, not necessarily one registered argmax label.

Experiment: OPTIC–1F: factored skill composition. Artifact identifier: DIAL-SkillOpt–1D v2.
Factorization: probe selection, uncertainty-qualified admission, and typed extension rendering.
Result: end-to-end accuracy improved from \(10/15\) to \(15/15\), and false extensions fell from \(3/15\) to zero.
Boundary: the hypotheses, probes, and renderer vocabulary remained registered; this supports robust composition under noise, not autonomous invention.