R0043/2026-04-01/Q002/SRC03/E01¶
SR 11-7 effective challenge and independent validation requirements
URL: https://www.federalreserve.gov/supervisionreg/srletters/sr1107.htm
Extract¶
SR 11-7 requires:
- Independent validation — "Validation involves a degree of independence from model development and use, and generally is done by staff who are not responsible for model development or use and do not have a stake in whether a model is determined to be valid."
- Effective challenge — "Models should be challenged by people who are smart, independent, and have enough influence to demand changes."
- The OCC has confirmed that SR 11-7 applies to all quantitative models, including ML and AI models.
JUDGMENT: SR 11-7's "effective challenge" is the financial services equivalent of anti-sycophancy: it requires that model outputs be independently questioned rather than accepted at face value. The parallel is structural — sycophancy in AI produces outputs that resist challenge; SR 11-7 mandates that outputs receive challenge regardless. However, SR 11-7 addresses the governance process (who challenges, how) rather than the model behavior (tendency to agree).
A sycophantic AI model used in banking would violate the spirit of SR 11-7 if its agreeable outputs reduced the effectiveness of the challenge function — but the regulation does not directly test for or prohibit sycophantic model behavior.
Relevance to Hypotheses¶
| Hypothesis | Relationship | Strength |
|---|---|---|
| H1 | Supports | Formal requirement exists addressing the phenomenon through governance |
| H2 | Partially contradicts | Requirements exist but do not name the model behavior |
| H3 | Supports | Addresses governance mechanism, not model behavior |
Context¶
SR 11-7 was written in 2011, long before LLM sycophancy was a concern. Its applicability to AI models has been confirmed by regulators, but its provisions were designed for traditional quantitative models (credit scoring, risk modeling) rather than generative AI. The "effective challenge" concept maps well to anti-sycophancy but was not designed for that purpose.