What stops sensitive data reaching the model in the first place?

Most architectures answer this with a filter that runs when a question is asked. That is a genuine control, but it operates after the sensitive material has been embedded into a retrievable index, so its effectiveness depends on the filter behaving correctly for every query shape anyone will ever attempt. The stronger answer is ordering: exclude the material before it becomes retrievable, so no formulation can reach it.

In Centralpoint, classification and redaction execute during the transformation that prepares a record for indexing. Data Cleaner applies the organization's own dictionary — its terms, its policies, its regulatory vocabulary, imported through Data Transfer — so identifiers are removed and sensitive categories recognized before the embedding layer receives anything. The approach is rule-driven rather than model-driven, which has a consequence worth stating plainly: the organization neither pays a model to perform the classification nor inherits the risk of a model performing it badly. 451 Research identified this ordering as the structural difference from platforms that filter after a model has already processed the data.