Model risk tiering: not every model gets the same scrutiny
Risk-based tiering done honestly — the tier decides the manifest, not the other way around.
Not every model deserves the same audit burden, and not every model can bear the same scrutiny. Risk-based tiering is how you say which is which — and the tier decides the manifest, not the other way around.
The honest version of model risk management is that not every model is high-risk, and treating a low-risk model as high-risk wastes scrutiny that the high-risk one needs. The EU AI Act codifies this as a legal taxonomy — unacceptable, high-risk, limited-risk, minimal-risk — with obligations that attach per tier. SR 11-7 (supervisory guidance) reaches for the same idea in bank terms: model risk is tiered by use, materiality, and complexity, and the controls scale with the tier. The tier is not a label you pick; it is a classification you defend, and the manifest you produce has to match the tier the model is in.
The tier decides the manifest
The manifest is not the same for every model. A high-risk system — one that affects credit, employment, essential services — gets the full per-output, hash-chained manifest: model version hash-pinned, input NodeVersions, agent trace, override, replay anchor. A limited-risk system gets transparency without the full chain; a minimal-risk system gets voluntary documentation. The tier decides the manifest, not the other way around — a firm does not pick the manifest and then argue the tier; it defends the tier and then produces the manifest the tier requires. That is what makes the tiering honest: the manifest is the consequence, not the cause.
Why tiering is honest, not convenient
Tiering done honestly is not a way to reduce burden; it is a way to focus it. A firm that classifies a credit model as minimal-risk to avoid the manifest has not reduced its burden; it has mis-classified, and the mis-classification is the finding an examiner leaves with. The honest version is that the tier is defended against the model’s actual use and impact, and the manifest is produced to match — which means a high-risk model carries the full chain whether the firm finds it convenient or not, and a minimal-risk model carries only what the tier requires. The tier is a classification you defend; the manifest is a deliverable you produce; and the first is the load-bearing one.
The tier decides the manifest, not the other way around. A firm defends the tier and then produces the manifest the tier requires — it does not pick the manifest and argue the tier.
The tiers and their force
The tiers carry different force. The EU AI Act’s risk tiers are binding law — the classification is a legal one, and the obligations are statutory. SR 11-7’s tiering is supervisory guidance — expected, not legislated, but no less load-bearing in a bank review. NIST AI RMF and ISO/IEC 42001 (voluntary) document risk-based processes but do not define the tiers; they layer governance on top of whichever binding or supervisory tier applies. Naming the force behind the tier is part of defending the tier — a firm that says "we tiered this model under SR 11-7" has said something different from "we tiered this under the EU AI Act," and an examiner reads them differently.
- EU AI Act tiers = binding law; obligations attach per tier and are statutory.
- SR 11-7 tiering = supervisory guidance; expected in a bank review, not legislated.
- The tier decides the manifest: high-risk gets the full chain; minimal-risk gets documentation only.
- Tiering is a classification you defend, not a label you pick to reduce burden.
What tiering does not exempt
Tiering does not exempt any model from the honesty the manifest carries. A minimal-risk model still does not certify truth; a high-risk model still does not certify truth. Certified ≠ True across all tiers — the manifest attests provenance and reproducibility, and the refusal to attest truth does not change with the tier. What changes is the depth of the record, not the kind of claim. Accuracy is a calibration loop gated by Gate 1 and Gate 1b, and we are pre-revenue, building the POC; the tiering is writable, the manifest fields exist, and the calibration that would back any tier’s accuracy claim is gated.
Tiering is not a way to reduce burden; it is a way to focus it. The honest version is that the tier is defended against the model’s actual use — and the manifest is the consequence, not the cause.
— model risk note