01
Classify
Every field in every system is assigned a sensitivity tier before ingestion: tier 1 for protected and regulated data, tier 2 for clients and counterparties, tier 3 for the seller's own operations. The tier decides how aggressive the next steps are.
02
Detect
Deterministic rules for structured identifiers, an open-weight model pass for free text, and a second pass that reads the surrounding context rather than the token alone. The union of the HIPAA eighteen and sixty-plus further classes: account and policy numbers, plates, VINs, device serials, docket numbers, biometric images, unique phrasings a search engine would resolve.
03
Replace
Entity-consistent pseudonyms: the same surrogate for the same person, client, payer or vendor in every source, so the corpus stays coherent and the twin stays faithful. Geography is generalized to state. Ages are banded. Account-type identifiers are dropped or replaced with consistent surrogates.
04
Shift
Every record belonging to one entity is shifted by the same random offset, so absolute dates are gone and every interval is exact. For health data this is permitted under Expert Determination, which is why we use it rather than Safe Harbor, whose year-only dates would destroy the timelines a lab is paying for.
05
Generalize
Rare events re-identify. A single very large claim in a small adjuster's book names the firm. The magnitude is bucketed and the workflow is kept, per dataset, tuned to the tier.
06
Gate
k-anonymity on quasi-identifiers with k set per tier; a re-identification red team with a model and a person; human QA on a sample; the exclusions check by hash. Any failure returns the dataset to generalization.
07
Document
Per dataset: method (Safe Harbor or Expert Determination), who signed, the identifier classes covered, the date-shift design, k, tiers, red-team result, watermark id. Delivered with the data.