Model Risk Management from SR 11-7 to SR 26-2

Between March and June 2023, the accuracy of GPT-4 at telling prime numbers from composites fell from 84% to 51%, with nothing on the customer's side having changed (Chen, Zaharia & Zou, 2023, How is ChatGPT's behavior changing over time?, arXiv:2307.09009). A bank that had validated that model in March would have been relying in June on a validation of something that no longer existed. Model risk management (MRM) is the discipline banks built for exactly this: a model can be wrong in its design, wrong for the use it is put to, or quietly different from the version someone checked.

For fifteen years the reference text was the Federal Reserve's SR 11-7 (April 2011), mirrored by OCC Bulletin 2011-12. On 17 April 2026 the Fed, OCC and FDIC replaced it with revised interagency guidance, issued by the Fed as SR 26-2 (Board of Governors, 2026, SR 26-2: Revised Guidance on Model Risk Management). As of September 2026 that is the operative US text, and it changed both tone and scope.

What carried over: effective challenge and three kinds of evidence

The idea that survived intact is effective challenge: "critical analysis conducted by objective experts who evaluate model risk and effect appropriate changes throughout the model lifecycle." It needs expertise, independence, and "the organizational standing and influence to effect any change." Drop any one and what remains is review theatre. A brilliant validator reporting to the business line that owns the model has expertise without independence; a committee that cannot block a launch has independence without influence.

Validation still breaks into the three components SR 11-7 set out:

• Conceptual soundness: are the design, assumptions and data choices defensible for the stated purpose? The 2026 text now says outright that for some models "interpretability measures or benchmarking to other models" may be more practical than examining the theory behind them.
• Outcomes analysis: do outputs match real-world outcomes, through back-testing, outlier analysis and holdout comparison?
• Ongoing monitoring: does the model still perform as expected as products, clients, data and markets shift?

Model risk combines inherent risk (assumptions, complexity, input quality, data constraints) with materiality, which is exposure (often measurable as portfolio size) plus purpose (regulatory and risk-management uses rank higher). Schematically, \(R_m = f(\text{inherent}) \times g(\text{exposure}, \text{purpose})\): rigour scales with \(R_m\), and an immaterial model may need only an inventory entry and monitoring.

What changed in 2026

Tone. SR 26-2 says it "does not set forth enforceable standards or prescriptive requirements; accordingly, non-compliance with this guidance will not result in supervisory criticism," though the underlying unsafe-or-unsound-practice authority remains. It is written mainly for banking organisations with more than $30 billion in total assets.

Definition. A model is now "a complex quantitative method, system, or approach that applies statistical, economic, or financial theories to process input data into quantitative estimates." Simple spreadsheet arithmetic and deterministic rule engines fall outside it.

Scope. A footnote carries the most consequential sentence: "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance." Traditional statistical models and "non-generative, non-agentic AI models" stay in scope. The OCC's companion bulletin says the agencies plan a separate request for information on AI in banking (OCC Bulletin 2026-13).

The UK went the other way. The PRA's SS1/23, effective 17 May 2024, explicitly covers the risks of AI and machine-learning techniques to the extent they are used in models (PRA, SS1/23 Model risk management principles for banks). So two major prudential regulators currently disagree about whether an LLM belongs inside the MRM perimeter at all.

Applying the discipline to an LLM system anyway

Being out of scope does not mean the risk is unmanaged. SR 26-2 itself says an organisation's broader governance "should guide the determination of appropriate governance and controls" for tools the guidance does not cover. In practice, the three validation components map across with some translation.

Conceptual soundness shifts from the weights to the system you control: prompt, retrieval corpus, tool permissions, guardrails, and fit between task and documented limits.

Outcomes analysis becomes evaluation against labelled cases, and the statistics matter. Suppose 400 graded outputs show a 5% error rate. A normal-approximation 95% interval is

so the defensible claim is "somewhere between about 3% and 7%." If the risk appetite is 4%, 400 cases cannot show that the system meets it.

Ongoing monitoring has to cover a failure SR 11-7 never had to think about: the provider changing the model behind a stable API name. Pinning model versions, keeping a fixed canary set, and re-running outcomes analysis on every provider release is the LLM version of back-testing.

When it breaks

Effective challenge without authority. The commonest failure is a validation function that writes findings the business can override by default. If a rejected model ships anyway with a "compensating control" nobody tests, the organisation has paid for challenge and received commentary.

Materiality judged by the model instead of the decision. A summarisation model looks harmless until its summaries feed a credit memo. Tiering has to follow where the output lands downstream.

Validation as a one-time gate. Budgets fund the pre-launch review and little afterward. For hosted models, that review can be stale within one provider release.

Principles-based can mean less. Supporters read SR 26-2 as sensible tailoring that ends checkbox validation of low-risk models. Critics worry that removing supervisory criticism for non-compliance, combined with the generative-AI carve-out, leaves the fastest-growing class of models with the weakest external pressure. Both readings fit the text, and which one proves right depends on what banks do voluntarily before the promised AI guidance arrives.