Material Change

A material change is a shift in methodology, training corpus, or target definition significant enough to require disclosure against a fixed prior state. A material change is only checkable if a sealed baseline captured before the change exists to compare the new state against.

A material change is a shift, in a model's methodology, its training corpus, its target definition, or its underlying data sources, significant enough that a counterparty relying on the model's output needs to know it occurred and needs to understand what specifically changed. Disclosing a material change honestly requires a fixed prior state to describe the change against, which is exactly the function a sealed baseline serves: without it, a disclosure becomes vague by necessity, describing only that updates were made, rather than specific, stating what population and target definition applied before the baseline and what applies after it. Institutions should define, in advance, what qualifies as material, a change in target definition, a meaningful expansion or contraction of the training population, a change in feature engineering that alters what evidence a model actually uses, rather than deciding case by case whether a given change rises to the level requiring disclosure, since a case-by-case standard is more easily bent by the pressure to avoid an uncomfortable disclosure in a specific instance. A material change disclosure that references a sealed baseline is checkable by a counterparty; one that does not is an assertion the counterparty is being asked to accept on the institution's own account of its history.