Proposal: residual re-identification risk report (the honest dual of detection)

[mazzasaverio]

Context

The vision's non-goals already say it plainly: pseudonymization "does not guarantee anonymization ... or immunity from contextual re-identification". Today that honesty lives in a docs page. The EDPB's 2026 anonymisation draft makes residual-risk assessment the legally relevant question, and the classic result (87% of the US population is unique on zip + birth date + gender) shows where pseudonymization actually fails: not on the identifiers it catches, but on the quasi-identifier combinations it leaves behind.

Every vendor claims "we redact". Nobody in the library space tells the caller what remains. Inverting the product question ("what did we remove?" → "what is still identifying?") is the same differentiation move as honest benchmarking (#62), applied to each individual output instead of the detector.

Proposal

A risk_report() on processing results, built under the existing safe-report constraint (no matched values, ever):

1. Quasi-identifier co-occurrence signals: counts and types only, e.g. "output still contains: 1 date, 1 location-like token, 1 age-like number; this combination is a known re-identification vector". Detection reuses existing detectors plus a small set of quasi-identifier classes (dates, ages, professions are out of scope for transformation but in scope for counting).
2. Below-threshold residue: detections that were found but filtered by policy (confidence, entity type) are already known to the engine; report their counts as residual risk instead of silently dropping the information.
3. Honest labeling as a hard requirement: the report is a heuristic signal, never a score pretending to be k-anonymity for free text. The docs page states what it can and cannot claim, in the same voice as docs/limitations.md.

Order-of-effects reasoning

• First order: compliance teams must assess residual risk; today they guess.
• Second order: as "we redact" becomes table stakes, procurement shifts to "show me what remains"; evidence beats claims.
• Third order: AI-incident liability and insurance markets need an underwritable artifact; a value-free risk report is that artifact.
• Fourth order: paired with processing receipts (Proposal: processing receipts (signed, value-free provenance manifests) #71), risk reports become the acceptance criterion for inter-organization data exchange.

🤖 Generated with Claude Code

https://claude.ai/code/session_012JiX3zWeEC28kmy5KmAXvf