AINumbers kernels are not SR 26-2 "models." The guidance defines a model as a method built on statistical, economic, or financial theory, and explicitly excludes deterministic, rule-based calculations from that definition. AINumbers kernels are deterministic, clause-cited recomputations, so most of them sit on the excluded side of that line. What a kernel can honestly be is one input to a bank's own outcomes analysis or benchmarking-to-other-models step, or, where a vendor model is in view, an independently-derived comparator number a bank's model-risk function can weigh against the vendor's output. This page maps kernel categories to the SR 26-2 clauses they can inform; it never claims the mapping does the bank's monitoring for it.
Supervisory Guidance on Model Risk Management, SR Letter 26-2, Attachment. Issued 2026-04-17 by the Board of Governors of the Federal Reserve System, the FDIC, and the OCC. Supersedes SR 11-7 (2011) and SR 21-8 (2021). Cover letter: federalreserve.gov/supervisionreg/srletters/SR2602.htm. Substantive attachment (the source of every clause below): SR2602a1.pdf, retrieved 2026-08-13.
The guidance defines a model as a method applying statistical, economic, or financial theory to produce quantitative estimates, and states the term excludes deterministic rule-based processes and software
with no such theory underpinning them. This is the clause that keeps AINumbers kernels outside the guidance's own definition of a model.
For a model a bank has already deemed low-materiality, the guidance sanctions a lighter duty: identify it and monitor its performance and the conditions under which it might become material, rather than full validation. A periodic, offline-checkable receipt is a plausible input to exactly that lighter duty.
Outcomes analysis, per the guidance, compares model outputs to corresponding real-world outcomes
to assess performance against objectives. Elsewhere the guidance notes that benchmarking to other models can be more practical than a full theoretical review for some models.
The guidance calls for ongoing monitoring and outcome analysis
to assess whether vendor models stay accurate and fit for purpose, while acknowledging a bank often cannot see a vendor's underlying code or methodology. An independently-derived, clause-cited comparator number is one way to partly close that gap.
This is a category mapping, not a per-kernel enumeration: there are hundreds of live kernels, and a hand-picked exhaustive list goes stale the moment a new one ships. These four categories cluster around the Part V and Part VII clauses above; a bank's own tooling can point at any AINumbers kernel's receipt, including ones beyond the examples linked here.
| Category | Example kernels | SR 26-2 clause it can evidence | How |
|---|---|---|---|
| Screening / AI-governance quality scorers | art-93art-97 | Part VII, vendor ongoing monitoring and outcome analysis | A scorer's cited-clause output on a bank's own screening-model tuning parameters is a dated, run-over-run data point. Drift in that score is the kind of thing Part V's ongoing model monitoring asks a bank to watch. |
| Deterministic rate / ratio / threshold recomputes | art-215art-218art-219art-220art-223art-224art-225 | Part V, benchmarking to other models and outcomes analysis | If a bank's pricing or loan-origination vendor model computes the same regulatory figure (APR, points-and-fees, a threshold flag), the equivalent kernel run on the same inputs produces an independently-derived comparator value. The bank's own benchmarking step does the comparing; the receipt is a raw input to it. |
| Capital / risk-ratio recomputes | art-180art-183art-184art-185art-254 | Part V, outcomes analysis and benchmarking | Same shape as the row above, applied to a bank's internal capital or IRRBB model rather than a consumer-lending pricing model. |
| Fair-lending disparity recomputes | art-229art-230 | Part V, outcomes analysis | A dated, independently-recomputed disparity figure a bank's fair-lending model-monitoring process can weigh against its own statistical model's output. The weighing is the bank's, not this page's. |
Every row above names a category of AINumbers kernel. Per the model definition in Part II, none of those kernels is itself an SR 26-2 model: the table names what a kernel's receipt can feed into, never what the kernel is. Nothing here claims coverage of a bank's model-risk needs; these are the categories that exist and fit, not a complete inventory.
Automated: ▶ Run the Fair Lending Disparity Audit Composer: computes HMDA rate spreads and disparate-impact metrics in one pass.
Automated: ▶ Run the Financial Crime Compliance Chain Composer: chains customer risk rating, FATF alignment, sanctions programme health, PEP and sanctions screening, fuzzy-match calibration, layering typology, AMLA typology scoring, anomaly detection, and structuring detection into an eleven-step financial-crime run.
Automated: ▶ Run the Insurer RBC Action Level Assessment Composer: chains NAIC RBC action-level calculation and claims straight-through-processing economics into a single guided session.
Automated: ▶ Run the IRRBB Supervisory Outlier Test Composer: calculates the IRRBB EVE shock, then evaluates supervisory outlier test EVE and NII.
Automated: ▶ Run the Life Illustration Self-Support Test Composer: runs the NAIC Model 582 self-support test, then computes the NAIC RBC action level.
Automated: ▶ Run the Mortgage Agency Pricing and Eligibility Composer: checks the conforming loan limit, then works an agency eligibility matrix and LLPA stack pricing.
Automated: ▶ Run the Mortgage APR Accuracy and Tolerance Cure Composer: solves the Reg Z Appendix J APR, then verifies TRID APR accuracy and fee-tolerance cures.
Automated: ▶ Run the Mortgage Compliance Preflight Composer: chains Reg Z threshold lookup, TRID APR verification, QM points-and-fees testing, and QM APR-APOR spread classification into a single guided session.
Automated: ▶ Run the Mortgage Government Loan Fit Composer: chains conforming loan limit, VA funding-fee and residual income, and FHA MIP eligibility checks into a single guided session.
Automated: ▶ Run the Fuzzy-Match Calibration Composer: scores fuzzy-match calibration and aggregates the agent-action audit trail.
Automated: ▶ Run the Sanctions Screening-Program Quality Composer: scores fuzzy-match calibration, checks screening list coverage, and scores the sanctions screening programme.
Automated: ▶ Run the Solvency II Reconciliation and Capital Composer: calculates the Solvency II SCR ratio, bridges SII to IFRS 17, and runs an insurance reporting readiness diagnostic.
v1.0 · August 2026 · SR 26-2 Attachment retrieved 2026-08-13
A category map from existing AINumbers kernels to the SR 26-2 clauses (Part V, Part VII) their output can inform when a bank's own model-risk function chooses to cite it, plus a companion tool (in progress) that re-verifies a batch of kernel receipts and assembles them into a dated bundle for that function to review.
Model validation, ongoing model monitoring, or outcomes analysis performed by AINumbers. The word "validated" is never used on this page to describe a kernel's own arithmetic: a kernel's output is checked against the cited clause, or independently recomputed; validation, in the SR 26-2 sense, is an activity of the recipient bank's own model-risk function.
An assurance grade. No pass/fail verdict on a bank's model, no opinion on a vendor model's fitness, and AINumbers is never named as the arbiter of anything a kernel touches.
Every artifact this mapping references ships at public-demo/example grade, stated here and on every generated bundle. It is not a substitute for a validation engagement.