5.3.1. Compliance Frameworks and Data Lineage

💡 First Principle: Regulatory compliance for GenAI requires proving not just that outputs meet standards today, but that you can demonstrate how any specific output was generated, what data informed it, and what controls were active at the time. This requires systematic data lineage and decision logging from the first line of architecture.

Model cards with SageMaker Model Cards:

# Create a model card for a fine-tuned deployment
sagemaker.create_model_card(
    ModelCardName='customer-support-fm-v2',
    Content=json.dumps({
        "model_overview": {
            "model_id": "customer-support-fm-v2",
            "base_model": "anthropic.claude-3-haiku",
            "fine_tuning_dataset": "s3://my-bucket/training/support-conversations-2024/",
            "intended_use": "Customer support question answering for Acme Corp products",
            "out_of_scope_uses": ["Medical advice", "Legal guidance", "Financial advice"],
            "model_version": "2.0"
        },
        "training_details": {
            "training_data_description": "100K anonymized support conversations from 2022-2024",
            "preprocessing_steps": ["PII removal", "Quality filtering >3/5 rating"],
            "training_date": "2024-10-15"
        },
        "evaluation_details": {
            "metrics": {"accuracy": 0.87, "hallucination_rate": 0.02, "safety_score": 0.98},
            "evaluation_dataset": "s3://my-bucket/eval/golden-support-queries/",
            "known_limitations": ["May underperform on technical hardware questions", 
                                   "Limited to English language"]
        },
        "responsible_ai": {
            "bias_evaluation": "Tested across customer demographics — no significant bias detected",
            "fairness_metrics": {"demographic_parity_diff": 0.03},
            "guardrails_applied": "arn:aws:bedrock:...:guardrail/GUARDRAILID"
        }
    }),
    ModelCardStatus='DRAFT'
)

Data lineage with AWS Glue Data Catalog: Every document that enters the RAG knowledge base should have a lineage record tracking its provenance:

# Register document lineage in Glue Data Catalog
glue.create_table(
    DatabaseName='genai-data-catalog',
    TableInput={
        'Name': 'knowledge-base-documents',
        'Description': 'Documents ingested into Bedrock Knowledge Base KBID12345',
        'Parameters': {
            'source_system': 'confluence-prod',
            'ingestion_job_id': ingestion_job_id,
            'ingestion_timestamp': datetime.utcnow().isoformat(),
            'data_classification': 'internal',
            'pii_scanned': 'true',
            'pii_findings': 'none',
            'knowledge_base_arn': 'arn:aws:bedrock:us-east-1:123456789:knowledge-base/KBID12345'
        }
    }
)

Metadata tagging for systematic attribution: Every AWS resource in the GenAI pipeline should carry standard tags for governance traceability:

GOVERNANCE_TAGS = [
    {'Key': 'data-classification', 'Value': 'internal'},
    {'Key': 'data-owner', 'Value': 'legal-team@company.com'},
    {'Key': 'compliance-framework', 'Value': 'sox-hipaa'},
    {'Key': 'pii-contains', 'Value': 'false'},
    {'Key': 'retention-policy', 'Value': '7-years'},
    {'Key': 'knowledge-base-id', 'Value': 'KBID12345'}
]

⚠️ Exam Trap: AWS Glue Data Catalog stores metadata about data assets (schema, location, lineage tags) — it does not store the actual data. When exam scenarios describe "tracking which documents were used to answer a specific FM query," that requires Bedrock Model Invocation Logs (which record the actual retrieved context in the request log), not Glue Data Catalog (which tracks data asset metadata).

Model cards and service cards: a model card belongs to each deployment, even one that uses an unmodified FM. The deploying organization documents its own intended use, out-of-scope uses, known limitations and evaluation results, because the provider's card describes the base model, not your HR screener. AWS AI Service Cards are AWS's published documentation of intended uses, limitations and responsible AI design choices for its AI services and Amazon models.

Data-subject requests: fine-tuned weights can encode training records. To honor an erasure request that covers fine-tuning data, remove the records from the source, retrain on the filtered dataset and delete the old model artifact. Lineage records tell you which models are affected. An access request must search every store that holds the person's data: source documents, the vector index, invocation logs, session and memory stores, and downstream databases.

AWS's side of compliance: Amazon Bedrock is in scope for SOC and ISO programs and is HIPAA eligible. Once a Business Associate Addendum (BAA) is in place, PHI can be processed under the shared responsibility model. Download AWS's audit reports, such as SOC 2 Type II, from AWS Artifact. AWS Config and Trusted Advisor check your own configuration but do not attest to AWS's controls.

Governance beyond your account: start retrofitted governance with an AI risk assessment covering use cases, affected users and potential harms. Then prioritize controls by severity. AWS controls govern only calls to your own Bedrock account. Keeping confidential data out of consumer AI services takes egress controls, such as blocking those domains at the corporate proxy, plus an acceptable use policy.

Reflection Question: A financial regulator requests an audit of all FM-generated investment recommendations made to clients over the past 12 months, including the exact data sources used to generate each recommendation. What combination of AWS services provides this audit trail, and what must have been enabled before the recommendations were generated?