Abstract
The previous paper in this series argued that closed-form analytics satisfy our substrate’s invariants for free, and that fitted models satisfy them only relative to a pinned artifact. This paper takes the other side. It describes Crucible, the extension that owns everything requiring a training loop — adaptive neuro-fuzzy systems fitted by automatic differentiation, random forests and decision trees fitted by classical optimization — and asks what it takes for a fitted model to be a legitimate participant in a governed decision rather than an opaque dependency. Our answer has two parts. First, the artifact must be a fact: a trained model enters the context with its dataset, its hyperparameters, its seed and its evaluation attached, so that which model decided this has an answer with the same provenance discipline as any other fact. Second, the pipeline that produces it should be a Formation and not a script — the stages of a training run are already Suggestor-shaped, and running them to a fixed point makes retraining a convergent process rather than a scheduled job. We have built the first and designed the second. The pipeline runs from a command line today and lifts into a Formation when a real retrain trigger pulls; we say so plainly rather than describing the design as though it were deployed.