Applicable Area

A model does not fail uniformly. It has a region where its predictions are trustworthy and a region where they are not, and the honest object of study is that region itself. That object is not a single number but a filtration of applicable areas, ordered by how much error one is willing to tolerate.

Applicability as a filtration

Fix a model and a tolerance. The set of inputs on which the model is trusted at that tolerance is one applicable area . Relaxing the tolerance can only grow it, so the family is nested,

This nesting is a filtration of rather than a hard in/out boundary. The target behaviour is honest out-of-distribution response. On the model should abstain or widen rather than extrapolate confidently. Applicability is therefore a coverage statement, namely what fraction of inputs the model will stand behind, and at what error.

Combining models by intersecting lax areas

An ensemble should be conservative, because each area is only laxly known, estimated and one-sided rather than exact. The applicable area of a combined model is modelled as the intersection of the members’ lax areas,

An input is inside the ensemble’s domain only where every member already stands behind it. This is deliberately pessimistic. Intersection shrinks coverage but raises the floor on trust, which is the correct trade when the cost of a confident-wrong prediction is high.

The split-free property

An applicable area is split-free when it cleanly separates an explainable pattern, meaning structure the model has captured, from genuine uncertainty arising from irreducible noise or true novelty. Where the split holds, low confidence can be attributed, distinguishing an unseen scaffold from a noisy assay rather than being lumped into one opaque score. Whether a given model admits a split-free area is not guaranteed, so it is a property to test for rather than to assume.

What this is meant to buy

The formulation is descriptive rather than a construction. It states what an applicable area is, namely a filtration; how areas compose, by lax intersection; and what good looks like, meaning split-free and honest under OOD and coverage. It does not say how to estimate from data. That is where the hedging lives, and where the empirical work has to happen.