From Policy Text to Traceable Classification

A plausible-looking hierarchy can still be unsupported. I therefore separated semantic extraction from structural construction.

The policy system uses a language model to extract classification records from source text and attach evidence references. Validation checks whether the extracted content remains grounded in the source. The hierarchy itself is then projected deterministically from accepted records rather than generated freely.

A separate metadata classifier starts from a rule catalog rather than allowing a model to invent labels. It retrieves candidate paths, applies deterministic context and domain hints, and uses constrained model selection only where needed. Once a path is selected, the classifier takes the security level stored in that rule instead of accepting a free-form level from the model.

The two systems were originally connected only through their shared taxonomy. I later added a file-based conversion layer that carries compatible extracted rules into the classifier while preserving source references and upstream review state.

For CS_RQ, a patient-field hint pushed the classifier toward a patient-information category and Level 3. The repository-provided expected path was an application-system category with Level 2.

The example matters because the error came from decision scaffolding rather than an unconstrained model output. Domain hints can improve automation while also creating systematic override errors.