11.6 From local repair to causal admission

Here integration is not synonymous with choosing a graph. A local field or bracket explanation must extend across intervention targets, overlapping charts, and intervention sequences without acquiring path-dependent contradictions. Approximate holonomy, closure under newly selected probes, and held-out multi-intervention prediction are possible integration observers. Only after this finite compatibility test and the registered controls should the method propose \(\Upsilon _{\mathrm{CLIC}}\), with a causal presentation map to a sketch containing a new latent variable, mechanism family, or factorization when applicable. The changed package must then make consequences not used to generate it.

. Bracket nonclosure is a localization signal, not a certificate of hidden confounding. DIAL-SKFM may infer a latent variable only when the intervention contract distinguishes it from missing edges, misspecified mechanisms, and other admissible explanations.

Likewise, selecting one graph from a supplied candidate class is causal model selection, not automatically scientific creativity. The creative DIAL-SKFM challenge begins when the system may introduce a previously unrepresented variable, mechanism family, latent dimension, or factorization and state what new interventions would distinguish it. Such a proposal is transformational relative to the old causal sketch, but it earns scientific value only after interventional admission.

CLIC thus fixes the object of construction—a causal mechanism or factorization—while leaving open how a system acquires the reusable skills needed to propose and test it. Chapter 12 moves to that second problem, replacing causal variables by typed skill programs while retaining the separation between local interaction, finite construction, and admission.