version: "1.0.1" name: agile-v-compliance description: Risk management, CAPA protocol, human gate approval records, AI agent security controls, and periodic revalidation. Load when running gates, auditing risks, handling CAPAs, or reviewing security posture. license: CC-BY-SA-4.0 metadata: version: "1.4" standard: "Agile V" compliance: "ISO 9001 6.1, ISO 13485 8.5, ISO 27001 A.5.23/A.8.3, 21 CFR Part 11, GxP/GAMP 5" author: agile-v.org sections_index: ["Risk Management", "CAPA Protocol", "Human Gate Approval Records", "AI Agent Security Controls", "Periodic Review & Revalidation"]

Instructions

Compliance protocols for Agile V. Requires agile-v-core loaded first.

Risk Management (ISO 9001 6.1 / AS9100D 8.1.1)

Append-only, cycle-tagged register in .agile-v/RISK_REGISTER.md: RISK-ID | Cycle | Category | Description | Likelihood | Impact | Severity | Mitigation | Owner | Status

Categories: Technical, Process, Compliance, Security. Severity matrix: High x High = Critical, High x Med = High, High x Low / Med x Med = Medium, rest = Low. Critical risks require Human resolution or documented acceptance before Gate 2.

When: Stage 1 = Req Architect identifies. Stage 2 = Logic Gatekeeper flags constraints. Stage 4 = Red Team finds residual. Cycle boundary = Compliance Auditor reviews.

CAPA Protocol (ISO 13485 8.5 / ISO 9001 10.1-10.2)

Triggers: CRITICAL finding, recurring NC across cycles, regression FAIL with no CR, 3-attempt escalation.

Record in .agile-v/CAPA_LOG.md: CAPA-XXXX with Cycle, Trigger, Nonconformity, Root Cause (5-Whys), Corrective Action, Preventive Action, Effectiveness Verification, Status (open -> corrective-complete -> preventive-complete -> verified-effective -> closed), Owner.

Workflow: Detect -> Record -> Analyze -> Correct -> Prevent -> Verify effectiveness. Compliance Auditor tracks open CAPAs at Gate 2, flags overdue (>2 cycles).

Human Gate Approval Records (21 CFR Part 11 / Annex 11)

Append-only in .agile-v/APPROVALS.md: GATE-XXXX with Gate type, Cycle, Scope, Decision (Approved/Conditional/Rejected), Conditions, Approver (full name), Role/Authority, Timestamp (ISO 8601), Signature Method, Evidence Reference (commit hash). Durable HITL (Phase 2): when closing a pending interrupt, include resume_token=[value] and INTERRUPT-ID=[INT-XXXX] matching .agile-v/CHECKPOINTS.md; Compliance Auditor verifies token pairing on Gate 2.

Rules: Name + role required (not just "Human"). Authority from matrix in config.json. Rejected = pipeline halts.

Checkpoint SLA: If CHECKPOINTS.md shows due_at passed with PENDING, escalate per project policy; append ESCALATED or EXPIRED row before forcing resume.

AI Agent Security Controls (ISO 27001 A.5.23 / A.8.3)

LLM Provider Registry in config.json: per provider record name, models, data_residency, retention, api_data_usage, approved_for classifications, review_date. Verify input classification vs provider approval before sending. Never send credentials/patient data unless provider approved. Least privilege per agent. Context sanitization on session end.

File Integrity: Git-tracked = verify clean status. Store hashes in STATE.md at Gates; verify before next stage. Flag unverifiable files to Human.

Periodic Review & Revalidation (GxP / GAMP 5)

Triggers: LLM model change, runtime/platform major update, skill file change, >5 CRs since last revalidation, 12-month interval.

Record in .agile-v/REVALIDATION_LOG.md: REVAL-XXXX with Date, Trigger, Scope, Results, Decision, Reviewer. Regression failure = new cycle trigger.

Model Tracking in config.json: model_versions with tier IDs + last_validated + validated_by. Any change triggers revalidation.