12.5 Evidence for skill factorization

Three empirical claims are carried by four archived protocols under the DIAL-SkillOpt name. The exact study is a held-out developmental calibration; the compound study is a sealed negative result; and the factored study is a separately registered diagnostic whose v1 and v2 records document one strict interface repair rather than two trials. Read together, they reveal why OPTIC must separate a skill program into typed obligations instead of asking one language-model edit to solve an entangled meta-learning problem.