regulated-ai-governance

Governance framework for AI agents and RAG systems in regulated environments — 41 examples, 25 jurisdictions, 10 AI frameworks, 2,631 tests.

Policy enforcement, audit logging, and compliance orchestration for AI systems that must satisfy FERPA, HIPAA, GDPR, EU AI Act, APPI, MAS FEAT, and 15+ additional regulatory frameworks.

Featured Article

MACF: The 6-Component Framework Every Enterprise Multi-Agent AI System Needs

A complete architectural walkthrough of the Multi-Agent Collaborative Framework (MACF), showing how regulated-ai-governance powers the ComplianceGate and PrivacyFilter components — with working Python examples, Mermaid architecture and sequence diagrams, and a full end-to-end request trace across all six framework components.

The problem this solves

AI agents in regulated environments can access and process data they are not authorized to see. Standard agent frameworks — CrewAI, AutoGen, LangChain, Semantic Kernel, Haystack — have no concept of regulated industry access control. When you deploy an agent in healthcare, financial services, or a government agency, the framework will not tell you whether the agent is allowed to access a record type, which decisions require human review, or how to produce the audit record the regulator requires.

This library provides those answers as composable Python filters: filters evaluate encoded rules against caller-supplied context. Context values are assertions until independently verified. Configure an audit sink to retain decisions; a permitted result does not establish legal compliance. See control evidence and audit guarantees.

Architecture

User / AI Agent Request
     │
     ▼
GovernanceOrchestrator.evaluate(context, document)
     │
     ├── Filter 1: Identity / Data Protection Gate ──────────────► FilterResult
     │      (FERPA, HIPAA, GDPR, APPI, PDPA...)
     │
     ├── Filter 2: Responsible AI Principles Gate ───────────────► FilterResult
     │      (METI, PDPC Model AI Framework, EU AI Act...)
     │
     ├── Filter 3: Sector-Specific Gate ─────────────────────────► FilterResult
     │      (MAS FEAT, MHLW, DORA, NIST AI RMF...)
     │
     └── Filter 4: Jurisdiction-Specific Gate ───────────────────► FilterResult
            (Cabinet Office, AHRC, DTA ADM, DSIT...)
                     │
                     ▼
              GovernanceReport
              overall_decision: APPROVED / DENIED / REQUIRES_HUMAN_REVIEW / REDACTED
              is_compliant: bool
              compliance_summary: str (human-readable for audit file)

Every filter is independently testable. Adding a new regulation means adding a new filter class — no existing filter is modified. Each context object is a @dataclass(frozen=True) — no mutable state passes through the governance pipeline.

Installation

pip install regulated-ai-governance

With AI framework extras:

pip install "regulated-ai-governance[crewai]"
pip install "regulated-ai-governance[langchain]"
pip install "regulated-ai-governance[autogen]"
pip install "regulated-ai-governance[llama-index]"
pip install "regulated-ai-governance[semantic-kernel]"
pip install "regulated-ai-governance[haystack]"

Quick start: CrewAI + FERPA

from regulated_ai_governance.integrations.crewai import EnterpriseActionGuard
from regulated_ai_governance.regulations.ferpa import make_ferpa_student_policy

guard = EnterpriseActionGuard(
    wrapped_tool=MyTranscriptTool(),
    policy=make_ferpa_student_policy(
        allowed_record_categories={"academic_record"}
    ),
    regulation="FERPA",
    actor_id="stu-alice",
    audit_sink=lambda rec: write_to_db(rec),
    block_on_escalation=True,
)
agent = Agent(tools=[guard], ...)  # drop-in replacement

Every evaluation emits a GovernanceAuditRecord — whether permitted or denied — with a regulation citation, suitable for a 34 CFR §99.32 disclosure log.

Quick start: Singapore Financial AI (MAS FEAT)

from regulated_ai_governance.examples import singapore_ai_governance as sg

context = sg.SingaporeAIContext(
    user_id="credit-model-v3",
    sector=sg.SingaporeSector.FINANCIAL_SERVICES,
    ai_risk_level=sg.SingaporeAIRiskLevel.HIGH,
    is_automated_decision=True,
    involves_personal_data=True,
    has_pdpa_consent=True,
    is_financial_institution=True,
    has_mas_approval=True,         # MAS FEAT Accountability A.2
    explainability_available=True, # MAS FEAT Transparency T.1
    bias_testing_done=True,        # MAS FEAT Fairness F.1
    human_review_available=True,   # PDPC Framework §2.3
    # ... other fields
)
orch = sg.SingaporeAIGovernanceOrchestrator()
results = orch.evaluate(context, document)
report = sg.SingaporeAIGovernanceReport(context=context, document=document, filter_results=results)
print(report.overall_decision)    # SingaporeAIDecision.APPROVED
print(report.compliance_summary)  # human-readable for audit file

Example catalog — 41 governance examples

Agentic AI Security (OWASP 2026 + NIST + ISO + MITRE + CSA)

| 39 | 39_owasp_agentic_top10_governance.py | OWASP Agentic AI Top 10 2026 | OWASPAgenticInjectionFilter (ASI01 goal hijack + ASI04 prompt injection defense), OWASPAgenticToolSafetyFilter (ASI02 tool misuse + ASI05 scope creep + ASI06 resource abuse), OWASPAgenticDataLeakageFilter (ASI03 memory poisoning + ASI07 insecure output handling), OWASPAgenticGovernanceFilter (ASI08 trust boundaries + ASI09 audit gaps + ASI10 unvalidated composition), 8 ecosystem wrappers (LangChain/CrewAI/AutoGen/SK/LlamaIndex/Haystack/DSPy/MAF) | | 40 | 40_governance_framework_auditor.py | Multi-Framework Governance Audit | Holistic governance gap-analysis — GovernanceFrameworkConfig (37 boolean controls), GovernanceFrameworkAuditor (37 controls spanning OWASP LLM Top 10 2025 (8), OWASP Agentic AI 2026 (4), NIST AI RMF + AI 600-1 (5), ISO/IEC 42001:2023 (9), MITRE ATLAS v5.1 (6), CSA ATF (5)), GovernanceAuditReport (scored 0–100, INITIAL/DEVELOPING/DEFINED/MANAGED/OPTIMIZING maturity, per-domain breakdown) | | 41 | 41_trilogy_security_audit.py | Unified Enterprise AI Security Audit | TrilogyAuditOrchestrator — combines RAG + Agentic + Governance auditors into a single enterprise assessment: TrilogySystemProfile (35 fields across all 3 auditor dimensions), combined_score = RAG×0.35 + Agent×0.35 + Gov×0.30, combined_maturity = minimum of three (weakest link), 7 cross-module gap correlations (XG-001 injection mismatch → XG-007 HITL gap across layers), TrilogyAuditResult with box-formatted enterprise security report |

Jurisdiction and regulation coverage matrix

AI framework integrations

Repository structure

src/regulated_ai_governance/
├── __init__.py               # Core exports
├── policy.py                 # ActionPolicy, GovernanceAuditRecord
├── regulations/
│   ├── ferpa.py              # FERPA 34 CFR §99
│   ├── hipaa.py              # HIPAA 45 CFR §164
│   ├── gdpr.py               # GDPR Art. 17/22
│   ├── glba.py               # GLBA 16 CFR §314
│   └── ...
└── integrations/
    ├── crewai.py             # EnterpriseActionGuard
    ├── langchain.py          # GovernanceCallbackHandler
    ├── autogen.py            # PolicyEnforcingAgent
    ├── llama_index.py        # ComplianceNodePostprocessor
    ├── semantic_kernel.py    # PolicyKernelPlugin
    ├── haystack.py           # FERPAMetadataFilter
    ├── dspy.py               # ComplianceModule
    └── maf.py                # GovernanceMAFMiddleware
examples/                     # 29 runnable governance examples (see catalog above)
tests/                        # 983 passing tests
docs/
└── ECOSYSTEM.md              # Full regulation and framework coverage matrix

Audit record format

Every filter evaluation produces a structured audit record:

{
    "filter_name":          "APPI_DATA_PROTECTION",
    "decision":             "DENIED",
    "reason":               "APPI Article 20-2: Opt-in consent required for sensitive personal information",
    "regulation_citation":  "Act on Protection of Personal Information (APPI) 2022 Article 20-2",
    "requires_logging":     True
}

JSON-serializable; suitable for compliance databases, SIEM ingestion, and regulator production requests.

Standalone Haystack Package

ferpa-haystack — FERPAMetadataFilter as a standalone Haystack 2.x custom component. Install with pip install ferpa-haystack. Adds FERPA identity-scope enforcement to any Haystack pipeline in two lines:

pipeline.add_component("ferpa_filter", FERPAMetadataFilter(student_id="stu_001"))
pipeline.connect("retriever.documents", "ferpa_filter.documents")

Google ADK Integration

ADKPolicyGuard — drop-in compliance layer for Google Agent Development Kit agents. Enforces FERPA, HIPAA, GDPR, GLBA, EU AI Act, and OWASP Agentic AI Top 10 2026 through ADK native before_model_callback, before_agent_callback, and before_tool_callback hooks. Supports BigQuery audit sink and multi-agent governance (Orchestrator → sub-agent authorization propagation).

from regulated_ai_governance.adapters.google_adk_adapter import ADKPolicyGuard, Regulation

guard = ADKPolicyGuard(regulations=[Regulation.FERPA, Regulation.HIPAA], agent_id="my-agent")
agent = LlmAgent(before_model_callback=guard.before_model_callback, ...)

Reliability and adoption

The source version and released distribution are separate artifacts. See CHANGELOG.md for unreleased changes; install a published version for reproducibility and record its version in deployment evidence.

• Audit guarantees and failure semantics
• Control provenance and applicability checklist
• Current roadmap

Contributing

Read CONTRIBUTING.md and GOVERNANCE.md. Run pytest tests/ -v before opening a pull request.

Citation

@software{rana2026rag,
  author  = {Rana, Ashutosh},
  title   = {regulated-ai-governance: Policy enforcement for AI agents in regulated environments},
  year    = {2026},
  version = {0.44.0},
  url     = {https://github.com/ashutoshrana/regulated-ai-governance},
  license = {MIT}
}

Part of the enterprise AI patterns trilogy

License

MIT — see LICENSE.