No model we run decides anything a regulator would ask about.
Last updated: September 24, 2026
Your AI and machine learning (ML) governance obligations now reach your vendors. Fannie Mae requires you to govern subcontractor and vendor use of AI/ML “no less protective” than your own, and both agencies can require you to disclose, on request, what AI is in use and what safeguards sit around it. That makes us part of your inventory, so here’s our half of it before you ask.
No AI/ML model produces, influences, or ranks any output that determines a regulated outcome.
Regulated outcome means any of: credit decisioning, pricing, valuation, adverse action, disclosure content or timing, or a servicing action. Everything we compute on that list comes from one deterministic engine, and there’s no model in it.
We enforce this in continuous integration (CI). We don’t just assert it. A gate walks the transitive import graph out from every module that owns or feeds a regulated outcome: the money math, pricing, fees, escrow, the compliance engine, the disclosure modules, the Truth in Lending Act and Real Estate Settlement Procedures Act Integrated Disclosure (TRID) tolerance, business-day math, automated underwriting, credit, valuation, Department of Veterans Affairs (VA) eligibility and funding-fee exemption, both loan origination system (LOS) payload serializers, and document filing, and it fails the build if a model endpoint ever becomes reachable from any of them. Not a warning. The build.
Three ways this kind of check usually fails quietly. Ours fails loudly instead, and we want it that way.
“Could not look” is not “clean”
If the gate can’t find a declared regulated module, that’s a failure, and the gate doesn’t skip it. Renaming or deleting a module can’t silently shrink the checked set.
A contract over nothing
An empty or shrunken module list fails. A reachability check over zero modules would otherwise pass while checking nothing at all.
A gate that has gone blind
A positive control runs inside the gate on every invocation, walking the surfaces that genuinely do reach a model. If it can’t find one there, the run fails, so a blind gate reports its blindness instead of a clean pass.
And what it can’t see, before you ask. It’s a static import reachability property. It doesn’t cover a dynamic import, a Hypertext Transfer Protocol (HTTP) call assembled at runtime, or a model reached indirectly through stored data. The gate prints that limitation on every run, and our governance disclosure lists it as a known limitation, so you don’t have to find it yourself.
Every model surface is optional, and none of them is in the regulated path.
These are the entries Apsis Line adds to your AI inventory. We don’t shrink your inventory, and we won’t claim to. We hand you each entry already tiered and scoped, with the evidence attached. Each one stays off until you configure it on its own, and enabling one doesn’t enable any other.
Loan-officer guideline assistant
Lender staff, behind the console sign-in · Tier 3, internal-only
Borrower question panel
Borrowers · Tier 2, borrower-facing, non-regulated
Staff co-pilot
Lender staff, behind the console sign-in · Tier 3, internal-only
Document intelligence
Classifies a borrower upload and extracts its fields; a staff signal in the shipped default · Tier 2
Retrieval first, and locally
On the question-answering surfaces, we run retrieval over a small bounded corpus before we call the model. The model receives only the retrieved passages plus whitelisted case facts. We ask it to phrase and cite them, and we don’t ask it to reason. We never ask it what the answer is.
Validation after generation
Our code validates every draft answer on those surfaces after the model returns, or a scripted refusal ships instead. An injected instruction can’t remove a citation requirement or introduce an unsourced figure, because the validation isn’t part of the prompt and nobody can prompt it away.
At the cap, the surface degrades instead of spending, so an overrun can’t reach the borrower
Each question-answering surface carries a per-tenant daily call cap and a throttle per internet protocol (IP) address. Over either one, the surface degrades to the deterministic retrieval answer and reports that it has done so. It doesn’t queue, and it doesn’t quietly keep spending.
One surface can’t compose a sentence of its own, and we keep it that way
We hand the staff co-pilot’s classifier a closed list of category identifiers, and it returns exactly one, or a sentinel meaning none of them. We treat anything outside the list as a refusal, so that surface can’t emit a sentence the model wrote. That’s a property of the code. It isn’t an instruction in a prompt.
Document intelligence reads the document, and says so
This is the one surface where the model receives borrower-supplied content instead of typed text: the uploaded document itself. We check ownership before we read a byte, we pin the answer to a closed list of document types and field keys, our code re-checks it, and if anything fails it doesn’t guess: it falls back to a result labeled simulated. In the shipped default it proposes a match and a person decides.
Your data doesn’t train anything we run, here or at a vendor
No model trains or fine-tunes on your data, your borrowers’ data, or ours. Our governance disclosure names our runtime model provider and exactly what each surface transmits to it. For the staff co-pilot’s relay, that’s staff-written text carrying no case facts at all. One surface, document intelligence, sends the uploaded document itself to that provider. We don’t fold that into the others: the disclosure names it on those terms. Disclosing our subprocessors upward is what makes your vendor-governance obligation dischargeable, so we put it in writing up front.
Development-time AI is a different question, and we answer it separately. AI-assisted software development is a software-lifecycle control. It isn’t an AI/ML system operating on borrower data. Generated code passes the same review gates, contract tests and CI enforcement as any other code, and no model participates at runtime just because it helped at build time. Reviewers most often conflate these two, so our disclosure separates them in a table before anyone asks.
Where these requirements actually reach a vendor
We want to get this exactly right, because a reviewer who checks will find it, and getting it wrong in either direction is the kind of error a risk team notices:
- Fannie Mae Lender Letter (LL) 2026-04 contains an operative clause. Its third bullet requires the seller/servicer to manage risks and governance of subcontractor and vendor use of AI/ML “no less protective of these requirements.” That’s quotable instrument text.
- Freddie Mac Guide §1302.8 does not contain a vendor clause. Its text reaches the Seller/Servicer only. The vendor reach comes from the scope of the bulletin that introduced the section. That’s a scope framing, and it isn’t section text.
We organize our own governance on the National Institute of Standards and Technology (NIST) AI Risk Management Framework (RMF), because both agencies use its vocabulary: mapping, measuring and managing AI risk, and the characteristics of trustworthy AI. NIST AI RMF is voluntary and non-certifiable, so that alignment is self-declared, and we say so. We don’t claim alignment with any framework we haven’t read.
Apsis Line is a technology vendor, and we aren’t a seller/servicer. Neither instrument binds Big Head Consulting directly. Nothing here is an attestation, a certification, an opinion of counsel, a substitute for your own third-party risk assessment, or a warranty of your compliance with any agency requirement. We wrote it to give you the factual record you need to discharge your own obligations with respect to us.
Our full disclosure states what we haven’t done. It carries a known-limitations section with eleven entries, including the one requirement we meet through the authority of a single officer instead of separately staffed ones, and the limit of the CI gate described above. A reviewer who finds an unstated gap discounts the whole document, so we hand you the list.
Send us your AI vendor questionnaire.
We send the full Vendor AI/ML Governance Disclosure to your risk team under a nondisclosure agreement (NDA). It covers the system inventory, the subprocessor list, the threat assessment, the bias position and the eleven known limitations. Most of it is already written.