Same Model, Two Rulebooks: Adjudicating One Package Under SR 11-7 and SR 26-2
Because our SR 11-7 and SR 26-2 domains run on the same reasoning engine — same graph topology, same deterministic tools, same challenge machinery, same proof layer — the same model package can be adjudicated under both regimes and the resulting dossiers diffed. The diff is clarifying. What is identical: the evidence record, every computed number, the structure of the challenge. What diverges: the classification (a Tier label vs. a materiality-scaled risk rating), the section anchors, the framing of findings, and the conclusion vocabulary. For institutions managing the transition, the diff answers the question every model owner is asking — what does the new guidance change for this model? — with an artifact instead of a memo.
A natural experiment in governance architecture
Regulatory transitions are usually analyzed in prose: law-firm client alerts, gap matrices, committee decks mapping old requirements to new principles. All useful, all abstract. The interagency replacement of SR 11-7 with SR 26-2 in April created the conditions for something more concrete. Because both regimes now exist on our platform as parallel domains sharing one engine — the regulation isolated in a citation pack, a classification scheme, and a conclusion vocabulary — you can hold the model package constant, run the adjudication twice, and read the difference between fifteen years of guidance as a document diff. Same evidence in. Same math. Two rulebooks. What follows is what the diff shows, using a representative validation case: an initial validation of a credit-decisioning model with a back-test, an outcomes review, and a handful of open documentation issues.
What the diff holds constant
The first finding is how much of the dossier does not move, and which parts those are. The evidence record is byte-identical: the same data extracts, the same documentation review, the same monitoring history. The deterministic tool blocks are byte-identical — the back-test statistics and math-verification results are computed in code from the same inputs, so they reproduce exactly under both regimes; a number cannot care which guidance is watching. And the structure of the effective challenge is the same: both runs steelman the model, attack in both directions, and force the conclusion to respond to surviving objections, because challenge machinery was never regime-specific — the agencies kept effective challenge across the transition for exactly that reason.
This is worth pausing on. Everything that constitutes epistemic work — gathering evidence, computing results, stress-testing conclusions — is invariant under the change of rulebook. If your validation platform requires rebuilding any of that to absorb a guidance update, the regulation was living in the wrong layer of your architecture.
What the diff shows moving
The classification. The SR 11-7 run locks a Tier 2 label from the legacy tiering scheme. The SR 26-2 run locks a model risk rating derived from inherent risk times materiality, where materiality combines exposure and purpose. Same model, same facts — different question asked of them. The practical consequence flows downstream: the 26-2 run’s challenge rigor and documentation depth are calibrated to the materiality rating, and the dossier says so explicitly, which is the proportionality the new guidance wants demonstrated rather than asserted.
The anchors. Every finding re-homes. What the 11-7 dossier anchors to the old guidance’s validation sections, the 26-2 dossier anchors to §IV (development and use), §V (validation and monitoring), §VI (governance and controls), §VII (third-party risk). The diff renders this as a clean mapping table — which sections absorbed which — and that table is itself a transition artifact most institutions are currently paying consultants to construct by hand.
The framing. The same open documentation issue reads differently under the two packs. Under 11-7’s requirements posture, it is a compliance gap against expected documentation. Under 26-2’s risk-based posture, it is a finding weighed against the model’s materiality — material enough to condition the determination, framed as sound-practice deviation rather than rule violation, with the non-prescriptive language of the guidance carried through. Neither framing changes the underlying fact; both dossiers cite the same evidence item. The diff makes the rhetorical layer of regulation visible — and shows it is only a layer.
The conclusion vocabulary. The 11-7 run discharges “validated with conditions.” The 26-2 run discharges “fit with conditions, stated at the model’s materiality” — and one legacy condition, calibrated to the old tiering, is re-derived against materiality in the new run. The determination is substantively concordant. The expression of it belongs to the regime.
What this is for
Three uses, in ascending order of importance. Transition management: a bank moving its inventory to the new guidance can run the diff per model and hand model owners a precise answer to “what changes for us” — the anchors move, the rating is recomputed, these two conditions re-derive, nothing else. Examiner conversation: a diffed pair is a powerful exhibit that the institution understands both where it was and where the guidance moved, with the invariant evidence record demonstrating that nothing substantive was gamed in the transition. And the architectural argument itself: the diff is empirical proof of the claim we have been making since April — that regulation, properly architected, is a configuration layer over a reasoning engine. The experiment only produces a clean diff because the engine and the rulebook were separable to begin with. Run the same experiment on a validation process where regulatory assumptions are smeared through prompts, templates, and tribal knowledge, and you cannot even define what to hold constant. The diff is not just a deliverable. It is the test of whether your governance architecture has layers at all.
Diff two dossiers and you can see, line by line, where regulation actually lives: not in the evidence, not in the math, not in the challenge — in the classification, the anchors, and the vocabulary of conclusion. Architecture that respects that boundary absorbs a regime change in a config. Architecture that doesn’t, rebuilds.