Agency-clear
Passes the formal program gates for VA IRRRL and FHA Streamline - the rules, not a guess.
Seasoning
Net tangible benefit
VA statutory recoupment
FHA combined-rate
Find the rate where a book of loans refinances itself.
Trigger Ladder pulls public rate data, models a synthetic pool of VA and FHA loans, and computes - for every candidate rate - how many files actually clear the real regulatory refinance tests. An AI writes the morning brief. An eval harness verifies every number before it ships.
By Brian Valentine - 34 years in mortgage lending, 12-state licensure.
~5,000
Modeled loans
VA + FHA / synthetic / SQLite
~17
Candidate rate rungs
scored top to bottom
3
Clearance layers
agency / economic / call
100%
Brief figures checked
verified before ship
The trigger ladder
One question, priced rung by rung.
For each candidate rate the desk re-scores the whole pool. Lower the available rate, and more of the book crosses into worth-refinancing. The live desk computes these from current public data - the ladder below is schematic.
Candidate rate / loans clearing each layer, out of ~5,000
Illustrative - not live market data
Agency-clear
Economically-clear
Call-clear
6.50%
6.25%
6.00%
5.75%
5.50%
5.25%
5.00%
4.75%
4.50%
Read it down: as the candidate rate falls, more of the pool clears each layer and the bars widen. A loan is refinance-ready only when it clears all three at once. Figures schematic - a rung is only real once it clears on live data.
Three clearance layers
Eligible, worth it, and doable are not the same test.
Most tools stop at eligibility. Trigger Ladder keeps three independent screens separate on purpose - a loan can pass one and fail another, and only files that clear all three are genuinely actionable.
Economically-clear
Genuinely worth doing under real break-even math - not just permitted on paper.
Break-even period
Cost recovery
Rate delta
Call-clear
The file is actually executable and saleable - nothing structural is blocking the close.
Executable
Saleable
Not blocked
Why keep them apart? Collapsing eligibility, economics, and executability into one score hides exactly where a deal dies. Separate layers make the failure legible.
How it works
Public data in, a verified brief out.
01
Pull public data
The 10-Year Treasury yield plus 30-year VA and FHA average mortgage rates.
source: FRED (Federal Reserve) + Freddie Mac PMMS
02
Model the pool
A synthetic set of roughly 5,000 VA and FHA loans, held in SQLite. No real borrower data.
03
Compute the ladder
Score the pool across about 17 candidate rate rungs against all three clearance layers.
04
Write the brief
The Anthropic API drafts a plain-English morning read on what the ladder is saying.
05
Verify before ship
An eval harness checks every number in the brief against the data. If a figure does not reconcile, the brief does not ship.
The eval harness
Every figure in the brief is checked against the data before a single word ships.
The AI writes the read, but it does not get the last word. Each number is reconciled against the computed ladder. A brief that cannot be verified is held back - so the language stays honest to the math.
Numbers reconciled
Claims traced to data
Fail-closed on mismatch
The honest part
What this is, plainly.
- An educational portfolio project on a modeled, synthetic loan pool - no real borrowers, no real book of business, no servicing.
- Public data only: FRED and Freddie Mac PMMS. Nothing proprietary.
- Any ladder or figure shown here is illustrative and schematic - not today's real market.
- Not financial advice. No NMLS-regulated activity takes place here.
The live desk
Run the ladder on live public data.
Open the Streamlit app to watch the rungs re-score and read a verified morning brief.