Your Rental-Income Checklist Expires November 1. Build the Five-Oppy Rule Diff Before It Lies.
By Anna with Oppy
Yesterday, Fannie Mae changed the map for qualifying rental income. Your team probably still has the old map in a shared drive, three loan officer cheat sheets, two onboarding decks, one branch manager’s head, and a prompt somebody wrote on a Friday.
On November 1, those artifacts do not become vintage. They become risk.
Fannie Mae’s September 2 Selling Guide announcement restructures rental-income policy, adds requirements for short-term rental income, standardizes treatment for certain investment properties purchased within 45 days of the subject property, creates a departing-residence framework, and strengthens lease validation. The rental-income changes are encouraged immediately and required for applications dated on or after November 1, 2026. The same announcement makes separate UAD 3.6 highest-and-best-use changes and NACLI expansion effective immediately.1
This is not a document problem. It is a compilation problem.
A policy PDF is written for careful human interpretation. A business runs on prompts, checklists, forms, scripts, CRM fields, calculators, training, quality-control tests, and decisions made while somebody eats lunch over a keyboard. The dangerous distance sits between the new sentence and the old workflow.
Build a Rule Diff: five narrow oppies that turn an authoritative policy change into a cited, tested, approved operating release. Not an AI underwriter. Not a legal oracle. A change-control line that makes stale instructions easier to find than to defend.
The 59-Day Problem
The mortgage market is not pausing while lenders update binders. MBA’s September 2 survey showed seasonally adjusted purchase applications rising 2% in one week, while adjustable-rate mortgages reached 8% of applications, their highest share in five weeks.2 More inventory and more product variation create more files with facts that resist the standard script.
Operators are also moving from dashboards toward conversational execution. On a mortgage podcast published September 2, Garrett Locklear offered the useful half of every prompting course:
“The best way to start learning how to get really good output is to focus on quality input. And quality doesn’t necessarily mean quantity.”3
Correct. The missing half is version control. A perfect prompt with a superseded policy is a confident way to be wrong.
Recent reporting points in the same direction. Inman’s Roxanne Hale argues that AI should organize professional expertise rather than invent it, and that people must retain the final read. RISMedia reported that a national brokerage COO’s mandate now includes aligning operational infrastructure with an AI-platform rollout. Mike DelPrete put the adoption problem more sharply while analyzing a different real estate product category:
“A defensive product can be announced in a day; the agent infrastructure behind it takes years.”6
Policy operations have the same physics. Publishing an update is fast. Finding every place the former rule still breathes is the work.
What Changed, and What Must Move
The official announcement belongs with qualified mortgage and compliance professionals. The table below is an impact map, not an underwriting interpretation.
| Policy area | Fannie Mae’s stated change | Operating surfaces to inspect |
|---|---|---|
| Short-term rental income | New eligibility, documentation, and qualifying-income treatment | Intake, document requests, worksheets, processor prompts, exception queues |
| Properties purchased within 45 days | Standardized rental-income requirements | Asset timelines, property matching, duplicate-address checks, underwriter checklists |
| Departing residences | Market-supported rents, reserves, and PITIA offset limitations replace reliance on leases alone | Conversion forms, reserve calculations, appraisal data, decision trees |
| Lease agreements | Minimum terms, validation, payment timing, and related-party restrictions | Review prompts, fraud flags, document-validation rules, escalation policies |
| UAD 3.6 highest and best use | Residential-use and reporting clarifications for UAD 3.6 assignments | Appraisal intake, URAR review, exception reasons, quality-control tests |
| NACLI | Expanded lender and transaction eligibility | Product matrices, tribal eligibility references, appraisal pathways, training |
A human can read this and nod. A production system must know which prompt, field, form, calculation, training clip, and test must change, who approves it, when it activates, and what happens to a file that began before the effective date.
That is the Rule Diff.
Build the Five-Oppy Rule Diff
Oppy lets teams create specialized AI employees, give them knowledge, define handoff rules, connect them to business tools, and assign narrow jobs.7 Do not use one giant agent with a heroic prompt. Use five small agents with boring boundaries. Boring is underrated. Air traffic control has very few jazz solos.
| Oppy | Narrow job | It may produce | It may not do |
|---|---|---|---|
| Source Clerk | Watch named authorities and preserve exact versions | Source packet, date, effective date, URL, checksum, owner | Substitute a secondary article for an available primary source |
| Clause Cutter | Compare approved old and new sources clause by clause | Structured change objects with exact citations | Decide borrower eligibility or fill a policy gap |
| Dependency Detective | Locate every operating artifact touched by a change | Ranked prompts, forms, fields, scripts, SOPs, training, and tests | Rewrite or publish an artifact |
| Shadow Underwriter | Run approved cases through old and proposed workflows | Pass, fail, divergence, citation, and unresolved-question reports | Approve a loan, contact a consumer, or alter production data |
| Release Clerk | Route questions, record approvals, schedule activation, preserve rollback | Signed release record, activation time, owner, rollback package | Activate without required human approval |
The boundaries matter more than the names. One agent finds. One extracts. One maps. One tests. One releases. Nobody quietly performs all five because it feels efficient.
Give the Clause Cutter a Contract, Not a Vibe
The fastest way to ruin policy automation is to request a “helpful summary.” Helpful summaries smooth ambiguity. Operations need visible ambiguity.
ROLE
You are the Clause Cutter for mortgage policy change control.
AUTHORITIES
Use only the approved source packet. Treat commentary, prior memory,
and secondary articles as non-authoritative.
TASK
Compare the approved prior version with the current version.
Create one change object for every added, removed, relocated,
clarified, or newly effective requirement.
REQUIRED FIELDS
source_name
source_url
source_version
page_and_heading
change_type
exact_old_text
exact_new_text
affected_population
transaction_type
documentation_named
calculation_or_threshold_named
effective_date
open_question
human_owner
RULES
Quote before paraphrasing.
Never infer an unstated eligibility decision.
Never turn an example into a universal rule.
Mark relocated text separately from substantively changed text.
If two authorities appear to conflict, stop and escalate.
If the effective date is missing, stop and escalate.
OUTPUT
A change-object table, then an unresolved-questions table.
Every row must include an exact source citation.
The Clause Cutter’s value is not literary elegance. It is refusing to hide the seam between what the authority says and what the team wishes it said.
Search for Behavior, Not File Names
Most change projects search document titles. That misses the dangerous copies.
The Dependency Detective should search for behavioral fingerprints: language from the former rule, old thresholds, old evidence lists, legacy calculator labels, canned processor emails, call scripts, quality-control reasons, and training recordings where a human explains the previous process.
For every match, it answers five questions: Where is it? Who owns it? Does it influence a consumer or credit decision? Can it run after November 1? What breaks if it is wrong?
Oppy’s public updates describe agents that learn from meeting transcripts, query connected knowledge, propose actions with confirmation, and report across communications, CRM activity, and pipeline data.8 That makes calls and internal conversations part of the dependency map. A policy can survive in software. It can also survive in the sentence a team leader repeats every Tuesday.
Test the Change Before the Change Tests You
The Shadow Underwriter works in a sealed environment with synthetic or properly de-identified cases approved by the organization. It compares the old workflow with the proposed workflow and explains each divergence with a source citation.
Do not ask it to prove the system is right. Ask it to find how the system could be wrong.
| Test | Question | Required control |
|---|---|---|
| Green path | Does an ordinary, documented case produce the approved checklist and route? | Proposed output matches the human-approved result |
| Red path | Does a plainly unsupported evidence pattern stop instead of improvising? | No calculation, no outreach, explicit escalation |
| Boundary | What happens on an application dated October 31 versus November 1? | Correct version selected from a controlled effective-date table |
| Contradiction | What if the lease, appraisal rent, and tax history disagree? | Conflict surfaced, not averaged into false certainty |
| Provenance | Can each document request and branch be traced to authority? | Exact page and heading on consequential output |
| Regression | Did the rental-income update alter an unrelated workflow? | Unaffected tests remain unchanged |
BCG recently described a finance insight agent that automated more than 90% of business-intelligence work in nine months, compressing a transformation that had previously taken three years.9 The transferable lesson is not the percentage. Narrow agents can remove extraction and packaging work while people retain interpretation and accountability.
Release on a Clock, With a Door Back Out
The Release Clerk treats November 1 as a version boundary, not a calendar reminder.
Every production artifact receives a rule-set version. Every application receives the approved version for its date and governing policy. Each changed artifact has a business owner, compliance approval, test evidence, activation time, and rollback package. Any artifact still invoking the former rule after activation is blocked and reported.
| Release field | Example |
|---|---|
| Rule set | FNMA-RI-2026-11-01 |
| Authority | SEL-2026-08 plus approved Selling Guide topics |
| Applies to | Applications dated on or after November 1, 2026 |
| Affected artifacts | Prompt IDs, form IDs, SOP sections, training clips, tests |
| Human approvals | Business owner, compliance, legal if required, model risk, operations |
| Test evidence | Passed set, failed set, accepted exceptions |
| Activation | Timestamp and deployment owner |
| Rollback | Prior version, reversion steps, incident contact |
The Rule Diff does not replace counsel, underwriting judgment, compliance review, or model governance. It gives those people a controlled surface to review instead of asking them to inspect an entire company by flashlight.
The Pattern Travels
A rule change rarely arrives labeled “AI problem.” It arrives as an insurance bulletin, a title-underwriting memo, an MLS rule, a property-management ordinance, a dental payer update, or a court filing that quietly changes what a script may say.
| Business | Change source | Rule Diff target |
|---|---|---|
| Insurance | Carrier or state guidance | Intake, quoting, renewal, notices, quality control |
| Title | Underwriting bulletin | Search review, commitment workflow, closer checklist |
| Brokerage | MLS or association rule | Listing intake, agent training, transaction review |
| Property management | Screening, fee, notice, or habitability rule | Leasing scripts, notices, resident communications, escalation |
| Dental office | Payer policy or state consent rule | Eligibility checks, estimates, outreach, records |
| Legal office | Court rule or filing requirement | Intake, deadline calculation, document assembly, attorney review |
The source changes. The control pattern does not: preserve, compare, map, test, approve.
Measure Whether the Rule Reached the Work
Do not count summaries. Count evidence that the new rule reached production.
| Metric | Definition |
|---|---|
| Detection latency | Publication to preserved source packet |
| Citation completeness | Change objects with exact source location |
| Dependency coverage | Known operating artifacts searched and classified |
| Human exception rate | Changes correctly escalated for interpretation |
| Shadow divergence | Proposed behavior that differs from the approved result |
| Stale-artifact count | Production items invoking the former rule after activation |
| Release lead time | Publication to approved operating release |
A zero escalation rate is not excellence. It is evidence that the agent learned to hide uncertainty. Reward clean stops.
The Shorter Letter
Fannie Mae changed rental-income policy on September 2. The mandatory date is November 1.1 Your problem is not reading eight pages. Your problem is finding every old sentence those eight pages just made dangerous.
Build five oppies. Give each a narrow job. Make them quote before they summarize, map before they modify, test before they release, and stop when authority ends.
A stale checklist rarely looks broken. It looks familiar. That is why it gets trusted.