The Layer Above AI

A language model predicts continuations over a context window. That is the whole mechanism. It holds no knowledge of a particular organization, forms no view about what it should be shown, retains nothing between requests and exercises no judgement about what it may say. Everything that makes it useful inside an enterprise arrives from outside it: which records enter the retrieval surface, which instructions load and in what precedence, which identity is asking, what budget applies, what is retained afterwards. That outer tier is where the organization's own intelligence lives, and it is the part that does not arrive with the model. Describing the model as the product and this tier as plumbing inverts the actual distribution of value, which is why organizations that invest only in model selection find the investment expires within a year.

Centralpoint is that tier by construction rather than by positioning. Classification and redaction execute at index time so the surface is bounded before any request exists; skills load in five fixed tiers with governance first and force-loadable; SkillTokenBudget bounds each execution; and the Interaction Log retains what was assembled. Swapping the model beneath changes none of it — which is the practical test of whether a platform sits above the model or merely in front of it.