Fair Lending AI Decision Logging
Fair Lending compliance for AI credit systems requires more than bias testing. Regulators and examiners expect institutions to produce verifiable records of how AI credit decisions were made, what data trained the model, and that decision records have not been altered after the fact.
CertifiedData provides the cryptographic infrastructure for Fair Lending auditability — tamper-evident records, reason codes, and certified training data provenance in a single verifiable chain.
What Fair Lending requires from AI credit systems
ECOA / Regulation B requires that credit decisions be based on permissible factors. When AI models are used, examiners expect institutions to demonstrate — not merely assert — that prohibited bases were not used in the model's inputs or training data.
Adverse action notices (required under ECOA and FCRA) must state the principal reasons for adverse decisions. AI systems must produce reason codes at the time of decision — not retroactively constructed from model internals.
CFPB and OCC supervisory expectations (including CFPB Circular 2022-03) indicate that unexplainable AI decisions are an ECOA compliance risk. Decision records must support examination-ready explanation.
Training data compliance — regulators expect institutions to verify that AI model training data was not derived from proxy variables for prohibited bases. Certified synthetic datasets prove this without disclosing sensitive data.
How CertifiedData addresses each Fair Lending requirement
Adverse action reason codes
Decision logs include a structured reason_codes array generated at decision time. Codes are locked into the tamper-evident hash chain — they cannot be modified after issuance. Enables automated adverse action notice production from cryptographically verified records.
Training data compliance proof
CertifiedData certifies synthetic datasets used for credit model training. The certificate records the generation algorithm, timestamp, and SHA-256 fingerprint. The certificate ID is embedded in every decision record — creating direct lineage from decision to training data.
Model version traceability
Every decision record includes the model ID and model version active at inference time. When a model is updated, the version change is captured in subsequent records — enabling regulators to identify which model version produced which decisions.
Tamper-evident audit trail
Decision records are append-only, chain-linked via SHA-256, and Ed25519-signed. Any modification of a record breaks the chain hash. Examiners can verify record integrity without requiring access to CertifiedData's systems.
Examination-ready documentation
The full decision record — including rationale summary, reason codes, model version, policy version, and certified dataset reference — can be exported as JSONL and presented to examiners as machine-readable evidence.
Sample credit denial record
A credit denial logged with CertifiedData includes all fields needed for adverse action notice production, regulatory examination, and consumer dispute response:
{
"actor": { "type": "agent", "id": "credit-scoring-model-v4.2" },
"decision": {
"label": "personal_loan_application",
"selectedOption": "denied",
"confidence": 0.94
},
"artifactReference": {
"certificateId": "cert_synthetic_credit_training_2024q3"
},
"explanation": {
"reasonCodes": [
"debt_to_income_exceeds_threshold",
"insufficient_credit_history_length",
"recent_derogatory_marks"
],
"rationaleSummary": "DTI 52% exceeds 40% policy threshold. Credit history 18mo < 24mo required."
},
"policy": {
"policyId": "consumer-lending-policy",
"policyVersion": "2024.09.15",
"policyHash": "a3f9c2..."
}
}The reason_codes are locked into the hash chain at decision time. The cert_synthetic_credit_training_2024q3 certificate proves the model was not trained on prohibited basis data.
Disparate impact analysis and HMDA reporting
Institutions conducting disparate impact analysis on AI credit decisions need reliable, complete decision records. CertifiedData's structured exports provide the decision-level data needed for statistical analysis — with verified record integrity so that the analysis rests on unaltered records.
Certified synthetic training datasets demonstrably exclude racial, national origin, sex, and other protected class indicators from model inputs — enabling institutions to document compliance with the disparate treatment standard.
Related
Decision Logging for Financial Services
Full financial services AI decision logging overview.
AI Decision Logging
How cryptographic decision logging works.
Synthetic Data Certification
Certifying training data to prove Fair Lending compliance.
AI Compliance System
Full regulatory compliance infrastructure for AI.