← Về thư mục
📄 / / profiles / r-and-d / raw_evidence / scientific-agent-skills / skills / treatment-plans / references / privacy_governance.md

Privacy and Data Governance

Last reviewed: 2026-07-23

No compliance claim

Passing a template or script does not establish de-identification, HIPAA compliance, authorization, lawful disclosure, security, or appropriate retention. Those determinations belong to qualified local privacy, security, legal, and records personnel.

Data classes

Use exactly one package-wide classification:

Never relabel real data as synthetic. Hashing, pseudonymization, redaction of obvious fields, using a patient code, or removing direct identifiers does not by itself make data de-identified.

HHS de-identification boundary

HHS describes two HIPAA Privacy Rule methods:

  1. Expert Determination — a qualified expert determines that re-identification risk is very small and documents methods and results.
  2. Safe Harbor — specified identifiers are removed and the covered entity lacks actual knowledge that remaining information could identify an individual.

This skill performs neither method. Free text, dates, geography, rare combinations, longitudinal patterns, and other contextual information can retain identification risk. HHS specifically notes that clinical narratives are information-rich and may allow identification.

For patient-derived material:

Do not place a de-identification analysis or patient-derived examples in this repository.

Minimum-necessary handling

HHS states that the HIPAA minimum necessary standard generally requires reasonable steps to limit uses, disclosures, and requests for protected health information, while identifying exceptions including disclosures to or requests by a healthcare provider for treatment.

This skill does not decide whether an exception applies. As a conservative process safeguard, always minimize what enters the package and follow the institution's current role-based access and disclosure policies.

Record:

Real-patient gate

Before any real-patient package is opened:

  1. Confirm local authorization and an approved environment.
  2. Confirm the accountable clinical and privacy owners.
  3. Confirm the minimum-necessary field set and intended recipient.
  4. Confirm no external service, model, API, search, telemetry, image, or cloud-processing step will receive content.
  5. Confirm content will not be copied into prompts, logs, examples, tests, screenshots, issue reports, or commit messages.
  6. Confirm retention, deletion, access, and incident-response rules.
  7. Run only local standard-library scripts against file paths.

If any confirmation is absent, use synthetic templates and stop before reading values.

Structured-data preference

Prefer discrete identifiers, enums, dates, booleans, role labels, and source references. Use bounded clinician-authored text only where exact transcription is necessary.

Avoid:

The generic subject_ref must be a locally controlled pseudonymous reference. It is not proof of de-identification.

Report minimization

Bundled scripts must not echo clinical values. Reports are limited to:

Do not run scripts with shell tracing. Do not redirect reports to shared logs. Review local command history policies before working with sensitive paths.

Local path controls

The scripts:

These are defense-in-depth controls, not privacy determinations.

Incident routing

If accidental disclosure, unauthorized access, or suspected mishandling occurs:

HHS OCR breach reporting sources are listed in source_ledger.md. The skill does not file reports.