annotating-variants
Annotates VCF variants and normalizes HGVS nomenclature with public, license-free annotators…
Generates and verifies 21 CFR Part 11-style audit trails — who/what/when, electronic signatures, and tamper-evidence — for OpenMed pipelines in GxP and clinical-trial (GCP) settings. Use when the user runs OpenMed in a regulated/validated environment and needs an attributable,
$ npx -y skills add maziyarpanahi/openmed --skill auditing-part11-trails --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/auditing-part11-trailsContext preview
The summary Claude sees to decide when to auto-load this skill.
Generates and verifies 21 CFR Part 11-style audit trails — who/what/when, electronic signatures, and tamper-evidence — for OpenMed pipelines in GxP and clinical-trial (GCP) settings. Use when the user runs OpenMed in a regulated/validated environment and needs an attributable,
name: auditing-part11-trails description: "Generates and verifies 21 CFR Part 11-style audit trails — who/what/when, electronic signatures, and tamper-evidence — for OpenMed pipelines in GxP and clinical-trial (GCP) settings. Use when the user runs OpenMed in a regulated/validated environment and needs an attributable, time-stamped, tamper-evident record of each processing action, electronic-signature manifestations, or computer-system-validation (CSV) evidence. Trigger keywords: 21 CFR Part 11, Part 11, audit trail, electronic signature, e-signature, GxP, GCP, GLP, GMP, CSV, computer system validation, data integrity, ALCOA, tamper-evident, contemporaneous. Pairs adjacent to OpenMed: maps directly onto OpenMed deidentify(audit=True) -> signed AuditReport with .sign(key)/.verify(key), whose repro_hash + HMAC give the tamper-evidence and attribution Part 11 expects. This is a compliance-enablement aid, not a validation certification." license: Apache-2.0 metadata: project: OpenMed category: compliance-regulatory pairs: adjacent version: "1.0"
In FDA-regulated **GxP** work (GCP clinical trials, GLP, GMP) any electronic record used to support a regulatory decision must meet **21 CFR Part 11**: it has to be **attributable** (who), **contemporaneous and time-stamped** (when), describe **what** changed, be **tamper-evident**, and — where a signing event occurs — carry a controlled **electronic signature**. These map onto the **ALCOA+** data-integrity expectations (Attributable, Legible, Contemporaneous, Original, Accurate, +Complete/Consistent/Enduring/Available).
OpenMed's `deidentify(..., audit=True)` already emits a deterministic, **PHI-free `AuditReport`** that you can `.sign()` (HMAC-SHA256) and later `.verify()`. That gives you the *tamper-evidence* and *attribution* primitives; this skill wraps them in the who/when/what/e-signature envelope Part 11 wants.
This is a **compliance-enablement aid**. Part 11 compliance also requires validated systems (CSV), SOPs, and access controls that live outside any single library — a QA/validation lead signs off.
attributable, tamper-evident record.
**e-signature manifestation** (meaning, signer, timestamp).
for an inspection.
| Part 11 expectation | 21 CFR cite | OpenMed mechanism | | --- | --- | --- | | Tamper-evident, accurate copies | 11.10(b),(c) | `AuditReport.to_json()` + `repro_hash` over the canonical payload | | Audit trail: what changed, when | 11.10(e) | `AuditReport.spans` (action per identifier), `input_hash`/`deidentified_text_hash`, `openmed_version`, `manifest_hash` | | Operational/authority checks; attribution | 11.10(d),(g) | `AuditSignature.key_id` (signer/key identity) + your envelope's user id | | Signature manifestation (name, date, meaning) | 11.50 | Your envelope fields `signer`, `signed_at`, `meaning` | | Signature/record linking, non-repudiation | 11.70, 11.200 | HMAC-SHA256 over the canonical payload via `.sign()` / `.verify()` |
The HMAC binds the signature to *that exact* report content: any later edit to a span, hash, or field changes `repro_hash`, so `.verify()` fails — that is the tamper-evidence.
import openmed, json, datetime as dt
note = "Subject S-014 (DOB 1962-08-09) reported headache on 2024-05-01."
# 1) Produce the deterministic, PHI-free audit record for this processing step.
report = openmed.deidentify(note, policy="hipaa_safe_harbor", audit=True)
# 2) Sign it with a controlled release key (stored in a vault / HSM, never in code).
report.sign(b"<release-hmac-key>", key_id="omv-signer-2026")
# 3) Wrap in a Part 11 envelope: who / when / what / signature meaning.
trail = {
"record": report.to_dict(), # tamper-evident, no PHI
"who": "j.smith@sponsor.example", # authenticated user (your IdP)
"when": dt.datetime.now(dt.timezone.utc).isoformat(),
"what": "PHI de-identification of source narrative (study X, subject S-014)",
"signature_manifestation": { # 21 CFR 11.50
"signer_printed_name": "Jane Smith",
"meaning": "reviewed and approved",
"signed_at": dt.datetime.now(dt.timezone.utc).isoformat(),
},
"system": {"openmed_version": report.openmed_version,
"manifest_hash": report.manifest_hash},
}
with open("part11_trail.json", "w") as fh:
json.dump(trail, fh, indent=2, sort_keys=True)
# 4) Later — verify integrity (optionally bind to the exact source/output text).
ok = report.verify(b"<release-hmac-key>", original_text=note)
assert ok, "AUDIT TRAIL TAMPERED OR KEY MISMATCH"1. **Authenticate the actor** in your own IdP/access system (Part 11 needs unique IDs and operational checks — outside the library). Capture the user id. 2. **Run the processing step with `audit=True`** to get the deterministic record. 3. **Sign** with a controlled release key from a vault/HSM; record `key_id`. Never embed the key in source or the trail. 4. **Build the Part 11 envelope** — who, when (UTC, contemporaneous), what, and the **signature manifestation** (printed name, meaning, timestamp) per 11.50. 5. **Append, never overwrite.** Store trails write-once (WORM / append-only store). The audit trail itself must be protected and retained. 6. **Verify on retrieval** with `.verify(key, original_text=..., deidentified_text=...)` to confirm neither the record nor the bound texts changed. 7. **Retain** per the study/retention schedule; keep keys and any reversible mapping in a separate, access-controlled store.
(`deident
Local-first healthcare AI: clinical NER and HIPAA PII de-identification on hardware you control. 2,200+ medical models, 35 model-backed PII languages, and Python, MLX, Android and browser runtimes. Apache-2.0 SDK.
Repo: maziyarpanahi/openmed
Annotates VCF variants and normalizes HGVS nomenclature with public, license-free annotators…
Choose the first OpenMed workflow skill for an intake, privacy, extraction, exchange, or…
Package multiple FHIR R4 resources produced from OpenMed output into a single valid…
Adversarially scan already-de-identified clinical text for residual identifiers and emit a…
Produce a signed, reproducible, no-PHI audit trail for an OpenMed de-identification run via…
Verify OpenMed de-identified output against all 18 HIPAA Safe Harbor identifier categories…