12.5.1 Exact mixed-witness calibration

12.5.1 Exact mixed-witness calibration

The exact calibration used linear worlds arranged in matched triplets: assimilation was sufficient in one world, accommodation was required in a second world with identical marginal diagnostics, and a third world withheld the mixed probe. GPT-OSS 20B edited only the decision-policy portion of an external Markdown skill, subject to typed-schema and false-extension gates.

This study is a held-out developmental calibration rather than a fully preregistered confirmation. The world split and admission truth were frozen before model calls, but the optimizer–executor interface was repaired using training and validation behavior before the final skill was hashed and the admission set was opened. On six locked admission worlds, the initial skill and ordinary SkillOpt without the mixed witness each solved \(4/6\). DIAL-SkillOpt with the witness solved \(6/6\), localized all six correctly, handled both incomplete-probe controls, and produced no false extensions. The two additional successes were exactly the accommodation worlds that marginal evidence could not distinguish from their assimilation partners.

Experiment: OPTIC–0: exact mixed witness. Artifact identifier: DIAL-SkillOpt–0.
Comparison: initial skill, SkillOpt without the mixed witness, and the complete mixed-witness skill.
Result: \(4/6\), \(4/6\), and \(6/6\) exact locked decisions, respectively, with zero false extensions.
Boundary: exact registered witnesses, fixed extension vocabulary, a deliberately small admission set, and pre-admission development of the executor interface.