← AI 101

Advanced training

Guidelines and Overlays Automation

How agency guidelines and GMFS overlays get ingested, structured, and surfaced as defensible answers at the point of work. The GMFS Guidelines & Overlays Engine — what it does, what it doesn't, and how to use it without misplacing your judgment.

19 minutes Builds on Module 4.1 Includes engine anatomy

Start here

What you'll be able to do after this lesson

Explain the pipeline

Describe — in plain terms — how agency guidelines and GMFS overlays get ingested, structured, kept current, and surfaced to the employee at the point of work.

Query in plain English

Ask the engine real questions in real language and get back a defensible answer with the right citation, effective date, and overlay applied.

Spot overlay-triggering changes

When a guideline updates, recognize whether the change should prompt an overlay review — and escalate it to the committee before the change ships through the system.

Three engine ideas

Ingestion, structure, surface

Ingestion

Source documents become structured content. Long agency PDFs from Fannie, Freddie, FHA, VA, and USDA, plus GMFS overlay policy, get parsed, normalized, and stored in a way the engine can reason over. The hard part isn't reading — it's keeping the store current.

Structure

Two layers, in a known relationship. The agency guideline says one thing; the GMFS overlay either matches it, restricts it, or adds a condition on top. The engine knows which layer applies and where they conflict.

Surface

How the engine actually answers a question. Plain English in, structured answer out — with the agency rule cited, the overlay layered, the effective dates checked, and a confidence band on top. The answer lands at the point of work, not in a separate research tab.

Trainer note: The value is not "AI can read guidelines." The value is a queryable, versioned, citable answer at the point of work. Without the citation and the effective date, the answer is a guess in a fluent voice — and a guess in a fluent voice is the dangerous failure mode.

Where the work happens

Three surfaces, one pipeline

The ingestion pipeline

Claude plus an agent flow that runs against updated source PDFs, normalizes new sections, flags conflicts with existing overlays, and queues anything ambiguous for committee review.

Behind the scenes

The structured store

The versioned, citable content the engine reasons over. Every answer the engine produces traces back here — and every change here has a date stamp and a committee link.

Source of truth

The query interface

Where the employee actually meets the engine — ELEVATE, a Claude chat session connected to the engine, or a dedicated tool. Same engine, multiple surfaces, one consistent answer format.

Point of work

Using and trusting the engine

Five rules to keep the engine honest

Always ask for the citation

"Cite the agency rule and any GMFS overlay with effective dates." Make the citation a non-optional part of every answer. An uncited engine answer is a guess in a confident voice.

Always check the citation against the source

Open the cited page, scan the cited section, confirm the language matches what the engine said. Two minutes of verification catches almost every engine failure.

Never publish an answer the engine isn't confident in

The confidence band is there for a reason. Medium and low confidence aren't "use anyway with hedging" — they're "escalate before relying on this."

Escalate ambiguous answers to the committee

If the answer addresses a near-question instead of your question, if the overlay-on-agency interaction looks wrong, or if the effective date is suspicious — the overlay committee is the right next stop.

Report any answer you can disprove

If you find an answer that's verifiably wrong, log it. Engine quality compounds with reported errors; without reporting, the same wrong answer ships to ten more people.

Prompt upgrade

Weak prompt

Can I do this loan?

Engine-ready prompt

Work-ready prompt

Borrower: 720 FICO, $80K W-2 income, no co-borrower. Property: $400K purchase in Florida, primary residence. Loan: $380K conventional, 95% LTV, 30-year fixed. Specific question: $8K gift funds from her biological parents — confirm any GMFS overlay on gift funds at 95% LTV. Cite the agency rule and the GMFS overlay, with effective dates for both.

Best patterns

Four engine workflows that earn the build

Pre-app eligibility check

Before the file goes in, run the scenario through the engine. Confirm the loan is eligible, identify any conditions, name the overlay considerations early — before they become rework.

In-flight condition resolution

Mid-process question — "is this gift funds documentation acceptable per the overlay?" — answered with the cited rule, in the conversation where the question came up.

Overlay impact analysis

When an agency publishes a change, the engine surfaces which GMFS overlays touch the changed sections — turning a multi-day manual review into a structured starting point.

Training content from current guidelines

Generate plain-English explanations of current rules and overlays for new hires — without the lag of waiting for the training team to rewrite anything.

Engine checks

Five things to verify on every engine answer

Citation present and accurate — open the cited section, confirm the language matches Effective date current — the cited rule is the version that's actually in force today The answer addresses the exact question — not a close-but-different one The overlay is applied correctly on top of the agency guideline — not in place of it, not skipped Edge cases route to a human — anything ambiguous, novel, or borderline goes to the committee

Employee rule: An engine answer with no citation is not an engine answer — it's a Claude paragraph wearing the engine's clothes. Verify the citation, confirm the effective date, and escalate anything the engine can't confidently cite. The cost of a fluent wrong answer in this domain is regulatory.

Drive the engine

Six exercises to build engine fluency

Use real scenarios from your pipeline. The point isn't to test the engine — it's to build your own habit of verification.

Open the engine anatomy in this lesson. Click each pin to see what every piece of a defensible answer should contain — and what you should verify.
Query the engine on three real loan scenarios from your week. Verify each citation against the source document.
Find one stale answer — an engine response that cites a rule that's been updated — and report it through the feedback channel.
Walk through one overlay-on-guideline interaction in detail. Note where the overlay restricts the agency rule and where it adds a condition.
Write one training paragraph generated from a current guideline. Verify every claim against the cited source before circulating.
Propose one new query pattern the engine should support — based on a question you've had to escalate that didn't have a clean engine answer.

Completion standard

You've finished this module when you can read an engine answer like an auditor — confirming the citation, the effective date, and the overlay-on-agency interaction in under two minutes, every time.