Model Harness

When an organization judges one model better than another for its work, it is usually comparing harnesses rather than weights. The harness decides which instructions are present, how retrieved material is selected and ordered, what the model is forbidden from doing, how tools are described, and what happens when the output fails a check. Two deployments of identical weights with different harnesses produce visibly different quality, and the difference is routinely attributed to the model. This misattribution is commercially convenient for providers, who bundle harness and weights so that the scaffolding's contribution appears to be a property of the model — and therefore something only that provider can supply. Once the two are separated, the weights become one component among several, competing on price, latency and locality like any other.

Centralpoint is a harness that outlives whichever weights it calls. Skills load in five fixed tiers ahead of any retrieved content; retrieval is bounded by classification applied at index time; consumption is budgeted per execution; and the assembly of each call is retained. Because the harness lives in the organization's own environment, improving it improves every model the organization will ever use, rather than improving one vendor's product.