Examiner-ready evidence mappings for the revised interagency MRM guidance, including the GenAI governance gap, third-party model risk, and concentration risk
GenAI Exclusion Notice
SR 26-2 explicitly states that "generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." A future request for information (RFI) addressing AI-specific model risk is planned but has not been issued.
What this means for your institution: Traditional quantitative models remain governed by SR 26-2. GenAI and agentic AI systems fall into a documented regulatory gap. Examiners are already asking about AI governance even though the formal guidance has not arrived. SWT3 provides tamper-evident witness records for both traditional MRM and AI governance activities, ensuring your institution is prepared regardless of when the RFI is published.
Who this is for: Bank examiners, Model Risk Management teams, Chief Risk Officers, fintech compliance officers, and internal audit professionals evaluating model governance frameworks. This crosswalk is designed for institutions that deploy quantitative models, AI/ML systems, or third-party vendor models for credit, fraud, pricing, capital, or operational decisions.
Today: When an examiner asks "how do you govern your AI models," your MRM team produces a spreadsheet of manual logs, screenshots of monitoring dashboards, and a memo from last quarter's committee meeting. The examiner has no way to verify when these artifacts were created, whether they have been modified since, or whether the committee meeting actually covered the model in question. The conversation becomes adversarial: the examiner asks for proof, your team scrambles to reconstruct it.
With SWT3: You query the witness ledger and produce a tamper-evident time series of governance anchors -- committee votes, validation sign-offs, monitoring results, back-test outcomes -- each independently verifiable via its cryptographic fingerprint. The examiner can confirm the evidence chain without trusting your infrastructure or your vendor's infrastructure. No screenshots. No reconstruction. No "we'll get back to you."
The shift: The examination moves from "prove you did this" to "let me verify what you already proved." Traditional MRM and AI governance produce the same evidence format, so when the GenAI RFI eventually arrives, your institution already has a retroactive evidence chain demonstrating that governance was in place before it was required.
| Field | Detail |
|---|---|
| Document | SR 26-2: Guidance on Model Risk Management (Revised) |
| Companion | OCC Bulletin 2026-13 (substantively identical) |
| Issuing Agencies | Federal Reserve, OCC, FDIC (joint interagency) |
| Effective Date | April 17, 2026 |
| Supersedes | SR 11-7, SR 21-8, OCC 2011-12, OCC 2021-19, OCC 1997-24, Comptroller's Handbook MRM |
| Key Change | GenAI/agentic AI explicitly excluded from scope; proportionality introduced; third-party model risk enhanced; non-binding status clarified |
| SWT3 Coverage | 279 procedures across 10 namespaces, 84 frameworks, 10 SDKs. FIN namespace (5 procedures) maps directly to SR 26-2 traditional MRM. AI namespace fills the GenAI governance gap. |
SR 26-2 is a comprehensive revision, not an incremental update. The following table summarizes every material change that affects model risk governance and SWT3 evidence collection.
| Area | SR 11-7 (2011) | SR 26-2 (2026) |
|---|---|---|
| Model definition | Broad: "quantitative method, system, or approach" | Narrowed: adds "complex" qualifier, excludes simple arithmetic and deterministic rule-based processes |
| GenAI / Agentic AI | Not addressed (did not exist at scale) | Explicitly excluded from scope; future RFI planned |
| Proportionality | One-size-fits-all expectations | Tailored to institution size, complexity, and model footprint; most directly applicable to institutions with >$30B in assets |
| Third-party models | Brief mention of vendor models | Enhanced requirements: vendor due diligence, ongoing monitoring, documentation of third-party model limitations |
| Concentration risk | Not addressed | New: institutions should assess single-vendor dependencies across model portfolios |
| Model inventory | Maintain an inventory | Enhanced: lifecycle tracking, tiering by materiality, linkage to business processes |
| Issuing agencies | Fed + OCC | Fed + OCC + FDIC (joint interagency) |
| Binding status | Supervisory guidance (implied binding) | Explicitly non-binding; non-compliance alone will not trigger supervisory criticism |
| Documents superseded | N/A | SR 11-7, SR 21-8, OCC 2011-12, OCC 2021-19, OCC 1997-24, Comptroller's Handbook MRM |
Practical impact: The narrowed model definition means fewer systems qualify as "models" under SR 26-2, but the systems that do qualify face more granular expectations around inventory, validation, and third-party oversight. Meanwhile, GenAI systems -- which arguably carry the highest model risk -- sit outside any formal guidance until the RFI is issued.
SR 26-2 applies to all banking organizations supervised by the Federal Reserve, OCC, and FDIC. However, the proportionality principle means the depth and formality of model risk management practices should reflect the institution's:
SR 26-2 defines a model as a complex quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates. The addition of "complex" excludes:
Critically, GenAI and agentic AI are not excluded because they are simple -- they are excluded because the agencies consider them too novel to govern under existing frameworks. This is a scope exclusion, not a risk exclusion.
SR 26-2 explicitly states that the guidance is non-binding and that non-compliance alone will not form the basis for supervisory criticism. However, examiners will still reference the guidance when evaluating MRM practices. Institutions that deviate materially from SR 26-2 expectations should document their rationale.
Institutions operating across multiple regulatory jurisdictions face overlapping model risk expectations -- SR 26-2 in the US, the EU AI Act in Europe, China's AI regulatory stack, and sector-specific frameworks like FDA AI/ML guidance for healthcare or India's DPDPA for data protection. SWT3 Witness Anchors are jurisdiction-neutral: the same evidence format satisfies multiple frameworks simultaneously through the bidirectional crosswalk engine (84 frameworks). An institution running models across US, EU, and APAC operations collects evidence once and maps it to each jurisdiction's requirements -- not three separate compliance workstreams producing three incompatible artifact sets.
The SWT3 FIN namespace defines 5 procedures that map directly to SR 26-2 traditional model risk management sections. Each procedure produces a tamper-evident Witness Anchor that records when a governance activity occurred, what the outcome was, and who performed it.
SR 26-2 Section III (Governance): The board and senior management should ensure model risk management is integrated into the overall risk framework with clear roles, responsibilities, and escalation paths.
What SWT3 witnesses: When the Model Risk Governance Committee votes to approve, modify, or retire a model, the vote is recorded with three evidence factors: quorum requirement, actual vote count, and approval decision. The Witness Anchor proves a properly constituted committee made a documented decision on a specific date.
Present the SWT3 Witness Anchor for FIN-GOV.1 alongside the committee charter. The examiner confirms quorum was met and the vote was cryptographically recorded. Cross-reference with meeting minutes for completeness. The anchor is independently verifiable with no vendor dependency.
SR 26-2 Section V (Model Inventory): Institutions should maintain a comprehensive inventory of all models, including version lineage, purpose, limitations, risk tier, and linkage to business processes. Lifecycle tracking from development through retirement is expected.
What SWT3 witnesses: Each time a model is deployed or updated, the SDK records whether the deployed model hash matches the approved version in inventory. PASS means production matches the registered version. FAIL means a mismatch was detected, creating an immutable record of when an unauthorized or untracked change occurred.
Filter the SWT3 ledger by FIN-MRM.1 for each model in your portfolio. A continuous chain of PASS verdicts demonstrates that production has always matched the approved inventory. Any FAIL creates an auditable record of when a discrepancy was detected and when it was resolved.
SR 26-2 Section VI (Validation): Validation should be conducted by qualified staff independent of model development. "Effective challenge" involves critical analysis by objective, informed parties who can identify model limitations and errors.
What SWT3 witnesses: When an independent validator signs off on a model review, the sign-off is recorded with evidence of who performed the validation, the sign-off status, and the number of days since the last validation for staleness tracking.
Present the FIN-VAL.1 Witness Anchor along with the validator's identity from ledger metadata. The examiner confirms: validation occurred, when it occurred, and who performed it. The staleness indicator enables comparison against your validation frequency policy. The record has not been altered since minting.
SR 26-2 Section VII (Ongoing Monitoring): Models should be subject to continuous monitoring to confirm they perform as expected. Deterioration in performance metrics should trigger review and potential remediation.
What SWT3 witnesses: The SDK records the performance metric (PSI, CSI, AUC, Gini, or institution-defined threshold) against its defined boundary at each monitoring interval. PASS means within bounds. FAIL means the threshold was breached. The timestamp and evidence factors create a tamper-evident time series of model health.
Present the time series of FIN-MON.1 anchors for each model. The examiner sees monitoring frequency, consecutive PASS verdicts showing stability, and if drift occurred, exactly when. The cryptographic integrity layer makes the entire monitoring history independently verifiable.
SR 26-2 Section VIII (Outcomes Analysis): Outcomes analysis compares model outputs to actual outcomes. This should be performed regularly with sufficient sample sizes to draw statistically meaningful conclusions.
What SWT3 witnesses: Each back-testing cycle is recorded with three evidence factors: required sample size, actual sample size, and whether results fell within tolerance. PASS means adequate data and predictions aligned with reality. FAIL means either insufficient samples or tolerance exceedance.
Present the FIN-OUT.1 Witness Anchor from the most recent back-testing cycle. Evidence factors prove sample adequacy and tolerance evaluation. The timestamp proves when the analysis was conducted. Compare frequency against your MRM policy to demonstrate adherence to your own standards.
SR 26-2 creates a documented regulatory gap that is unprecedented in federal banking supervision. GenAI and agentic AI are acknowledged as presenting novel risks, yet they are explicitly placed outside the scope of any active guidance. The planned RFI has no published timeline.
This gap is not theoretical. Financial institutions are already deploying GenAI for customer service, document summarization, code generation, fraud narrative drafting, and credit memo preparation. Examiners are already asking questions about AI governance during safety and soundness examinations even though no formal expectations exist.
Institutions should not wait for the RFI. The following SWT3 procedures provide witness evidence for GenAI governance activities that close the gap between what SR 26-2 covers and what examiners expect.
| Procedure | What It Witnesses | Gap Addressed | SR 26-2 Equivalent |
|---|---|---|---|
AI-DRIFT.1 | Output drift detected or confirmed absent at monitoring interval | GenAI performance monitoring | Section VII (Ongoing Monitoring) |
AI-MDL.5 | Model weights hash verified, version lineage recorded | GenAI model inventory and integrity | Section V (Model Inventory) |
AI-CHAIN.1 | Multi-model pipeline dependencies documented and traced | Model interconnection risk | No parallel in SR 26-2 |
AI-SUPPLY.1 | Third-party model provenance verified, vendor attestation recorded | Vendor AI supply chain governance | New third-party requirements |
AI-GOV.1 | AI governance policy reviewed and approved by oversight body | Board-level AI oversight | Section III (Governance) |
AI-EXPL.1 | Explainability method applied, output interpretable to stakeholders | Model transparency for effective challenge | Section VI (Validation) |
AI-HITL.1 | Human-in-the-loop review performed on high-risk decision | Human oversight for automated decisions | No parallel in SR 26-2 |
AI-GRD.1 | Safety guardrails active and tested | GenAI output controls | No parallel in SR 26-2 |
When the RFI is eventually issued and formal GenAI guidance follows, institutions that have been collecting SWT3 Witness Anchors will possess a retroactive evidence chain proving that governance was in place before it was required. This is a fundamentally different posture from scrambling to comply after the fact.
SWT3 produces the same tamper-evident Witness Anchors for AI procedures as for traditional FIN procedures. An institution presents a unified evidence chain covering both traditional models (FIN-GOV.1, FIN-MRM.1, FIN-VAL.1, FIN-MON.1, FIN-OUT.1) and AI models (AI-GOV.1, AI-DRIFT.1, AI-EXPL.1, AI-HITL.1, AI-GRD.1) to demonstrate comprehensive governance regardless of the regulatory gap.
Without this: Traditional MRM evidence lives in committee minutes and spreadsheets. AI governance evidence lives in separate tooling, separate formats, separate review processes. When the examiner asks for a portfolio-level governance view, your team stitches together incompatible artifacts. With this: One ledger, one verification method, one examiner conversation -- whether the model is a logistic regression or a 70-billion-parameter LLM.
Each of the following AI-namespace procedures maps to a governance activity that SR 26-2 expects for traditional models but does not address for GenAI. These are the procedures that bridge the gap.
Gap filled: SR 26-2 Section VII expects ongoing performance monitoring for traditional models, but GenAI outputs are non-deterministic and resist conventional statistical monitoring. AI-DRIFT.1 records whether output behavior has shifted beyond a defined threshold at each monitoring interval.
What SWT3 witnesses: The monitoring metric, its current value, and the threshold boundary. PASS means the model is behaving within expected parameters. FAIL means drift was detected. The time series of anchors creates a complete behavioral history for each GenAI deployment.
Present the AI-DRIFT.1 anchor series for each GenAI model. The examiner sees monitoring frequency and whether behavioral drift has been detected. This is the GenAI equivalent of the PSI/CSI monitoring that SR 26-2 expects for traditional models.
Gap filled: SR 26-2 Section III expects board-level governance for traditional model risk. For GenAI, the governance structure may differ -- an AI ethics committee, a responsible AI board, or an expanded MRM committee. AI-GOV.1 records that an oversight body reviewed and approved the institution's AI governance policy.
What SWT3 witnesses: The policy review event, the approving body, and the approval outcome. The anchor proves that AI governance is not ad hoc but follows a documented, board-sanctioned process.
Present AI-GOV.1 alongside FIN-GOV.1. Together, they demonstrate that both traditional MRM and AI governance receive formal oversight. The examiner can compare governance cadence between the two domains.
Gap filled: SR 26-2 Section VI expects "effective challenge" of models, which requires understanding how a model reaches its conclusions. GenAI models are inherently opaque. AI-EXPL.1 records that an explainability method (SHAP, LIME, attention visualization, or other technique) was applied and produced interpretable output.
What SWT3 witnesses: The explainability method used and whether the output was deemed interpretable by a qualified reviewer. This creates evidence that the institution attempted effective challenge even for models that resist traditional validation approaches.
Present AI-EXPL.1 anchors alongside the explainability artifacts (feature importance plots, attention maps, or narrative explanations). The anchor proves the analysis was performed; the artifacts provide the substance.
Gap filled: SR 26-2 does not address human oversight for automated AI decisions. As GenAI is deployed in credit memo drafting, fraud alert triage, and customer communication, the question of when a human must review AI output becomes critical. AI-HITL.1 records that a human reviewed a high-risk AI decision before it was acted upon.
What SWT3 witnesses: The review event, confirming that a qualified human evaluated the AI output. The anchor proves the institution maintains human oversight over consequential AI-assisted decisions.
Present AI-HITL.1 anchors for high-risk use cases (credit decisions, BSA/AML alerts, customer-facing communications). The examiner sees that human review is documented and timestamped, not assumed.
Gap filled: Traditional models do not generate free-text output, so guardrails are a concept unique to GenAI. AI-GRD.1 records that safety guardrails (content filters, prompt injection defenses, output validators) are active and have been tested.
What SWT3 witnesses: The guardrail configuration and test results. PASS means all guardrails are active and passed testing. FAIL means a guardrail was inactive or failed a test, creating a record that triggers remediation.
Present AI-GRD.1 anchors alongside guardrail test results. The examiner confirms that the institution is actively testing GenAI safety boundaries, not relying on vendor defaults.
SR 26-2 significantly expands expectations around third-party model risk. Where SR 11-7 offered only brief mention of vendor models, the revised guidance requires institutions to:
What SWT3 witnesses: Each time a third-party model is onboarded, updated, or re-evaluated, the SDK records the vendor identity, model version, and provenance attestation. The Witness Anchor creates a tamper-evident record of when the institution verified the vendor's model and what version was approved for use.
This directly addresses SR 26-2's expectation that institutions document and monitor vendor model usage. When a vendor releases a model update, a new AI-SUPPLY.1 anchor records whether the update was reviewed and approved before deployment.
What SWT3 witnesses: When models feed into other models -- a common pattern in fraud detection, credit decisioning, and risk aggregation -- the SDK records the dependency chain. Each anchor documents which upstream models contribute to downstream decisions, creating a traceable lineage that examiners can follow.
SR 26-2 does not explicitly address model interconnection risk, but examiners increasingly ask how institutions manage cascading model failures. AI-CHAIN.1 provides the evidence trail.
SR 26-2 acknowledges that institutions may not have full visibility into third-party model internals. This creates a validation challenge: how does an institution perform "effective challenge" on a model it cannot inspect?
SWT3 addresses this through behavioral monitoring rather than structural inspection:
This approach aligns with SR 26-2's proportionality principle: the institution demonstrates governance commensurate with the level of access it has to the model. For institutions deploying open-weight models (where weights are inspectable but the training process is not), the governance challenge is different -- see the Llama Stack compliance witnessing guide and the open-weight governance gap analysis for detailed procedure mappings.
SR 26-2 introduces concentration risk as a new consideration for model risk management. Institutions should assess whether over-reliance on a single model vendor, a single modeling methodology, or a single data source creates systemic vulnerability.
Many financial institutions rely on a small number of vendors for critical models -- credit scoring, BSA/AML screening, market risk valuation. If a single vendor's model fails, is recalled, or becomes unavailable, the impact can be institution-wide. SR 26-2 expects institutions to document these dependencies and maintain contingency plans.
SWT3 provides several evidence mechanisms for concentration risk governance:
Together, these procedures create an evidence chain that demonstrates the institution is aware of its vendor concentration profile and is actively monitoring for associated risks.
The following scenarios illustrate how SWT3 evidence addresses concentration risk questions that examiners may raise under SR 26-2:
| Scenario | Examiner Concern | SWT3 Evidence |
|---|---|---|
| Single vendor provides credit scoring, fraud detection, and BSA/AML screening | Vendor disruption affects multiple critical functions simultaneously | AI-SUPPLY.1 anchors grouped by vendor reveal the dependency. FIN-MRM.1 inventory shows which business lines are affected. |
| All GenAI use cases rely on one foundation model provider | Provider outage, pricing change, or policy shift affects all AI capabilities | AI-SUPPLY.1 anchors show vendor distribution. AI-CHAIN.1 traces which downstream processes depend on the single provider. |
| Vendor model is updated without institution review | Unauthorized model change in production | AI-SUPPLY.1 FAIL anchor records the moment a vendor model changed without approval. FIN-MRM.1 detects hash mismatch. |
| Model validation relies on vendor-provided test data | Insufficient independent validation | FIN-VAL.1 anchors document who performed validation and whether it was independent of the vendor. |
| Examiner Question | Where to Find Evidence |
|---|---|
| "Show me your model governance structure" | FIN-GOV.1 anchors. Each represents a committee vote with quorum status and approval decision. AI-GOV.1 anchors extend this to AI-specific oversight bodies. |
| "Is this model in your inventory?" | FIN-MRM.1 anchors for traditional models, AI-MDL.5 for AI/ML models. PASS means production matches the registered version. |
| "Who validated this model and when?" | FIN-VAL.1 anchors with validator identity in ledger metadata. The staleness indicator shows days since last validation. |
| "How do you monitor model performance?" | FIN-MON.1 anchor frequency for traditional models. AI-DRIFT.1 for AI models. Daily anchors demonstrate daily monitoring. |
| "What third-party models do you use?" | AI-SUPPLY.1 anchors. Each records vendor identity, model version, and provenance attestation for every third-party model in the portfolio. |
| "Do you have single-vendor concentration?" | Query AI-SUPPLY.1 and FIN-MRM.1 ledger by vendor. The distribution of anchors across vendors reveals concentration profile. |
| "How do you govern GenAI systems?" | AI-GOV.1 (policy), AI-GRD.1 (guardrails), AI-HITL.1 (human oversight), AI-EXPL.1 (transparency). These fill the SR 26-2 GenAI exclusion gap. |
| "Has this model drifted?" | FIN-MON.1 or AI-DRIFT.1 evidence factors. PASS means within threshold. FAIL means drift was detected and recorded. |
| "When was the last back-test?" | FIN-OUT.1 most recent anchor timestamp and evidence factors for sample adequacy. |
| "Can I verify these records independently?" | Yes. The public verifier or offline SHA-256 formula. No vendor dependency required for verification. |
| "How do I know records have not been altered?" | Each Witness Anchor is a SHA-256 fingerprint. A periodic integrity rollup binds all anchors into a tamper-evident digest. Any alteration breaks the chain. |
| "Where is the factor data stored?" | Clearing Level 2+: evidence factors remain on-premises. Only the cryptographic fingerprint traverses TeNova infrastructure. |
When documenting SWT3 in your Model Risk Management Policy or System Security Plan, consider language similar to the following:
Install the SWT3 SDK in your language of choice and begin recording model risk management evidence in minutes. All 10 SDKs produce cross-language compatible Witness Anchors.
pip install swt3-ai
from swt3_ai import Witness
witness = Witness(
api_key="axm_your_key_here",
endpoint="https://sovereign.tenova.io/api/v1/witness"
)
# Record a model governance vote
witness.witness("FIN-GOV.1", verdict="PASS",
factor_a="quorum_required:5",
factor_b="votes_cast:7",
factor_c="decision:approved"
)
# Record GenAI drift monitoring
witness.witness("AI-DRIFT.1", verdict="PASS",
factor_a="metric:psi",
factor_b="value:0.08",
factor_c="threshold:0.25"
)npm install @tenova/swt3-ai
import { Witness } from '@tenova/swt3-ai';
const witness = new Witness({
apiKey: 'axm_your_key_here',
endpoint: 'https://sovereign.tenova.io/api/v1/witness'
});
// Record a third-party model provenance check
await witness.witness('AI-SUPPLY.1', {
verdict: 'PASS',
factorA: 'vendor:acme_analytics',
factorB: 'model_version:3.2.1',
factorC: 'provenance:verified'
});Free tier includes 30-day anchor retention, Clearing Level 0-1, and full SDK access. View SDK documentation for all 10 languages.