A three-layer architecture for reliable, auditable AI decisioning
Written by Ari Gross
The first two pieces in this series made a pair of arguments: a generative model cannot be its own answer key, and a technology’s seat in the system should be chosen by asking whether you can afford if it fails. This one puts the conclusions to work.
If an LLM shouldn’t be the decider, and can’t be its own checker, then what does the right system actually look like?
It looks less like a single brilliant model and more like a well-run institution: a core that follows the rules identically every time, a reviewer who looks over its shoulder, and a clear escalation path for the cases the rules never anticipated. Three layers, each doing the one job it is built for.
Layer one: a deterministic core that decides
At the center sits the system of record — structured extraction, validation logic, and a rules engine that turns verified facts into decisions. Its defining property is that it is predictable and replicable in exactly the way regulators love: the same file produces the same decision, today or three years from now, and the reasons are written into the logic rather than reconstructed after the fact.
Two design commitments make this layer carry the compliance load on its own when it comes to mortgage AI systems:
- Every decision writes its own evidence. Not a log bolted on afterward, but a decision record produced as a byproduct of deciding: the inputs as received, the checks that ran, the rules that fired, the versions of everything involved, and the outcome. When a repurchase demand or a fair-lending exam reopens a file, the answer isn’t an investigation — it’s a lookup. Adverse action reasons fall out of the same structure, faithful by construction, because the rule that denied the file is the reason the file was denied.
- Everything is versioned, so the past stays reproducible. Rules change; guidelines change. A defensible mortgage AI system can always answer “what would this file have decided under the logic in force at the time?” — which is the question examiners actually ask.
Notice what this layer deliberately excludes: any component whose behavior can drift. That exclusion is what makes the next layer safe to add.
Layer two: LLM modules that verify
Wrapped around the core is generative AI in the seat the second piece argued for — the reviewer’s. LLM modules read the same documents the core processed and ask the questions rules can’t enumerate:
- Does this pay stub look like pay stubs look?
- Is the story this file tells internally consistent?
- Does anything here resemble the alterations we’ve seen before?
The critical design decision is not what these modules can do, but rather, what they cannot. Their output is confined to a prescribed, well-defined channel — a flag, a pointer to the evidence, and a plain-language note for the human who will look. They have no direct path to the decision itself; the constraint lives in the software’s permission structure, not in a policy document. That confinement is what converts an unpredictable technology into a safe one: the verifier’s worst day produces an unnecessary review, never a wrong decision.
The verifier isn’t checking the model’s work against the model — it’s checking a different system’s work against the source documents. Independence is what makes verification mean something.
Layer three: a disciplined path for the boundary
Every lending and insurance operation has a long tail: the self-employed borrower with four income streams, the handwritten supplement, the claim that fits no template, the document format no one has seen before. These boundary cases are where pure rules engines historically failed and where the temptation to “just let the mortgage AI decide” is strongest.
The architecture resists that temptation with routing. Every file carries a measured confidence that is derived from how the checks actually came out, not from asking a model how it feels, and confidence determines the path. High-confidence files, where answers can be verified, complete straight through on the deterministic core. Files the verifier flags, or that fall outside the core’s competence, route to the boundary path, where the LLM does its most valuable work: not deciding the hard case, but preparing and advising for it — reading the unstructured material, assembling the relevant evidence, laying out what’s unusual and why, making advisory recommendations — so that the human who decides starts from a briefing instead of a pile.
The result is that people see only the files that genuinely need judgment, and see them well-prepared. The rare case gets more expertise than it would in a manual shop; the routine case gets none, because it needs none.
Why this scales and compounds
Enterprise scalability falls out of the routing. Volume growth lands almost entirely on the deterministic path, where marginal cost is close to zero; the expensive tiers grow only with genuine ambiguity, not with volume. And the system improves in a way a static one can’t: every boundary case a human resolves is a candidate for additional system logic, and these updated logic rules moves a slice of tomorrow’s volume from the expensive path to the cheap, deterministic one. The core grows more capable; the boundary shrinks; the economics improve with age.
So does the defense. Each processed file adds to a body of validation records, decision histories, and exam artifacts that no competitor can replicate by copying the design. The architecture is the starting advantage. The accumulated evidence is the durable one.
The question to keep asking
These three blogs reduce to three questions worth carrying into any mortgage AI decisioning evaluation: Checked against what? Can we afford when it fails? And now the architectural one — when a file arrives that the system has never seen before, who decides it, and what can they prove afterward?
If the answer is “the model figures it out,” look closer. If the answer is “the core decides what it can prove, the reviewer flags what it can’t, and a well-briefed human decides the rest,” that’s a system built not just to work, but to take the stand.
