12.5.2 The compound-skill failure

12.5.2 The compound-skill failure

The preregistered compound-skill study introduced noisy finite evidence, active probe selection under cost, thresholded hypothesis admission, and selection among three typed extensions. One bounded Markdown edit was required to coordinate all three operations. Neither proposed edit passed every validation gate while strictly improving the registered selection vector. The procedure terminated as specified, froze no skill, and never opened the locked admission set.

This is a useful negative result. It does not show that the model lacked every required capability. It shows that a monolithic external skill and one scalar selection event obscured which obligation failed. OPTIC therefore treats factorization as part of the skill theory, not merely as prompt engineering.

Experiment: OPTIC–1: compound-skill termination. Artifact identifier: DIAL-SkillOpt–1.
Task: jointly learn active probing, noisy admission, and typed extension construction through one bounded edit.
Result: no admissible improving skill; the confirmatory run terminated before admission.
Epistemic status: registered negative result; no skill was frozen and the locked admission worlds remained unopened. The result localizes failure to the compound-skill contract rather than establishing model incapacity.
Lesson: failure localization requires persistent subskill traces and separate type and validation contracts.