Summary
Checking an invoice against an earlier remittance for a possible duplicate depends on relationships: same supplier, same account, overlapping service period. Blanket redaction replaces every supplier and account with [REDACTED], destroying exactly the information the task needs — the privacy control defeats the question. The requirement is therefore to decide per inference whether data may leave the trust boundary at all and, if so, in what form, with any transformation preserving the relationships the task depends on.
Opus offers three arrangements — inference inside the boundary, standard external inference, and external inference with Smart Aliasing — compared across data exposure, task quality, cost drivers and capacity constraints, explicitly as mechanisms to assess rather than a universal ranking.
Smart Aliasing works in stages inside the boundary: inputs are classified by modality, language, complexity, sensitivity and machine-readability; documents are converted to structured text preserving order, table membership, header associations and key-value pairs (necessary because an alias must be consistent across pages and repeated references); entities are detected and given a semantic type and role, so the same raw string can be an invoice date on one line and a service date on another; and a stable semantic alias service replaces resolved entities with typed, deterministic aliases whose mapping never leaves the boundary.
The external model therefore sees recurring SUPPLIER_01 and ACCOUNT_3 links that support a potential-duplicate flag, with two limits named: opaque aliases do not encode chronology, and surrounding document context still reaches the model. Responses are restored inside the boundary under policy, with validation checks blocking re-materialization of never-issued, corrupted or invalid alias combinations.
The closing trade-off is unusually direct: this is a privacy transformation, not formal anonymization — context can identify, detection can miss entities, resolution errors can wrongly link them — and the measurements that would settle its value are how much raw data still leaves, how much quality is retained, and how often restoration errs.