Abstract
A great deal of what organizations call machine learning is a mean, a variance, a linear fit and a threshold. This paper describes Prism, the closed-form analytics extension of our convergence substrate: eleven packs — descriptive statistics, z-score anomalies, linear regression, logistic classification, Gaussian naive Bayes, k-means segmentation, exponential smoothing, moving-average trend detection, weighted multi-criteria ranking, pairwise similarity and fuzzy inference — exposed as Suggestors that propose into a governed context. None of them has a training loop, and we argue that this is a substrate property rather than a modesty. A closed-form pack is a pure function of the facts it reads, so it satisfies the determinism, idempotency and commutativity axioms of the substrate for free; a fitted model satisfies them only relative to a pinned artifact, which is a strictly weaker statement and one that has to be maintained. That difference is exactly where we draw the boundary with the extension that does fit models. We also report a small piece of engineering we have come to rely on: the authority boundary between analytics and promotion is enforced by tests that are required not to compile, so that an attempt to give an analytic pack promotion power fails at build time rather than in review.