Solutions — instrument
Model risk management, for models that act.
Supervisory Guidance on Model Risk Management. Written in 2011 for models that produce numbers, and now read against systems that take actions — which is where the inventory question gets interesting.
SR 11-7 / OCC 2011-12
Federal Reserve · OCC
clause map
Status
In force
Applies to
US banking organisations using models in decisioning
What the platform evidences
- Model inventory
- Every model an agent used, with the column that earns the page: source. A runtime reporting the model that served a call is evidence; a human typing a name into a form is a claim. Every regime in scope asks for an inventory, and only the first kind is one.
- Configuration provenance
- Versioned snapshots of what each agent was at a point in time, so a grant issued against v1 reports drift when the agent becomes v2 rather than silently continuing.
- Ongoing monitoring
- Behavioural drift with an explanation attached — an agent whose behaviour moved with a governance act behind it is housekeeping; one that moved with nothing explaining it is the only row on the screen that is news.
- Effective challenge
- The record of who reviewed, what they refused, and what they approved — including the second named approver on authority above the regime's threshold.
Validation is not ours. SR 11-7's core demand is independent validation by people with the standing to challenge. RotaGrant records the validation status of a model and who set it. It does not validate a model, and a vendor claiming to satisfy SR 11-7 by shipping software has misunderstood the guidance.