A lending world model, built one piece at a time. Each piece below is
one drawing and one idea.
In one line. Do not make an AI decide every loan by hand.
Make it write the rulebook, check the rulebook, then let cheap rules do the work.
New to this, start at section A. In a hurry,
skip to the game ↓.
BLUE structure, reference, annotationORANGE the rule that failedALL FIGURES DRAWN FROM RECORDED RUNS IN /V3/
A · PLAIN WORDS
Three words you need
No finance degree required.
Terms of the game.
A loanA bank gives you money for a house and you pay it
back monthly, for years. The bank's job: guess whether you will pay it back.
An agentHere it means an automated decision maker: a piece
of software that says approve or deny on each loan. It can be plain rules, or a model.
The castBorrower wants the loan
→ broker finds them and passes it on
→ lender says yes or no
→ investor buys the loan. Risk flows down this chain.
The whole question: who or what should make the yes/no call, and can you
defend the call afterwards.
B · THE CAST
Every player is an agent
Four players in the chain. Each is its own agent, an automated decision maker
with its own goal. None of them is a person; all of them are software making calls.
Borrower agent
decides: apply, and how much to reveal
Wants: the loan, at the lowest rate.
Trick: some fudge the truth to look safer.
Wins approved cheap · Loses rejected, or cannot pay
Broker agent
decides: pass this lead on, and to whom
Wants: a commission per loan passed along.
Tension: paid to push deals, including shaky ones.
Wins deals close · Loses pushes fraud, gets clawed back
Lender agent
decides: approve, deny, refer
Wants: fund good loans, reject bad ones.
Bind: too strict loses business, too loose eats defaults.
Wins grows safely · Loses goes broke, or breaks rules
Bank and investor agent
decides: how much credit, and whether to repurchase
Wants: steady returns, no late surprises.
Power: a loan built on a lie gets sent back.
Wins loans pay off · Loses a whole vintage sours at once
Watch them play.Notice: a loan flows down the chain, then one defaults and everyone reacts.
Four agent types. Each one can run on any of four brains.
C · THE BRAINS
Same agent, four ways to decide
This is the experiment: four player types × four brains.
01 · Plain rules
0 tokens per decision
A fixed rulebook. Free, instant, fully readable. Never breaks a rule,
and never improvises.
02 · Model picks, rules guard
462 to 470 tokens per decision
The model chooses each move, but only from legal options. Flexible and
unable to cheat. Costs a little on every decision, forever.
03 · Model writes the rules RECOMMENDED
12,040 tokens once, then 0
The model writes the rulebook once, we check it against held-out loans,
then plain rules run it for free. 0% wrong on the measured 40,000 loan run.
04 · Model decides freely
251,360,000 tokens across 40,000 loans
No rails. Most flexible, most expensive, and it breaks the rulebook
outright: 100% of decisions wrong on the measured run.
The game runs all four player types, and you can put any of them on any of
these four brains.
STEP 1 OF 7
See the decision
Build an applicant yourself, drag the two dials.
Four brains decide, live
You decide who the applicant is.Notice: the rules give the same call every time; the unconstrained model
declines to commit on the same person.
Same applicant, four brains, four answers. Which one could you defend to
an examiner.
Below: the recorded reasoning for two brains on one file, drawn from
/v3/brains.json. Every step is shown, including the last one.
Same loan, two recorded traces.Notice: the rules cite a limit at every gate; the free-running model repeats
one step, then approves in violation.
If you cannot read how it decided, you cannot defend it or regulate it.
What the steps mean, in plain words
Income, the two-year average of pay, feeds the affordability check: does the payment
eat too much of that income, which lenders call DTI. Then reserves: do they hold
savings as a cushion. Then the decision. These are the checks a human underwriter
runs; the rules just run them for nothing.
One loan is cheap. Forty thousand is the test. ↓
STEP 2 OF 7
Cost it at scale
The rule planner spends 0 tokens, so it looks like the cheapest option.
It is not. Drag the loan count
Cost at 40,000 loans.Notice: the per-loan model columns scale with volume; the generated policy
stays flat, and human review time is the line that actually dominates.
Free just moves the cost. A model of the work is only honest once it
charges for the humans.
So far this is a quiz. Real loans do not resolve at the closing table. ↓
STEP 3 OF 7
Give it a cycle
Nobody sets a boom or a bust. The agents act, and the weather emerges.
All four agent types, over six years.
All four agents, one world.Notice: borrower demand and lender health rise together in the boom, and because
they are chained, they fall together too. Each line normalized to its own maximum.
Your turn: drag the lending dial.Notice: nobody dialed a boom. Loosen credit and the bubble inflates itself;
the looser you lend, the harder the fall.
A decision is not real until it has a delayed, coupled consequence, and
every agent feels every other agent's.
Zoom all the way in. What does one loan's whole life look like. ↓
STEP 4 OF 7
Follow one loan
A broker pushes a shaky file through. Who is holding it when it blows up.
Read to the last row
The life of a defective loan.Notice: it performs for 18 months, then the loss flows back up the chain.
Once a defect gets put back and the commission clawed back, the incentives
bite the agent that created them.
Now the real question: can a model write rules this good. ↓
STEP 5 OF 7
Generate the policy
Hand the model the published guide and say: write the rulebook.
Make it write a fresh one a few times
Free-write against tune-and-validate.Notice: every free-written rulebook scores near the floor; constrain the model to
tuning parameters we then check, and it reaches 100% every time.
The win is not more model. It is generate, then validate, then run for free.
STEP 6 OF 7
The real policies
Real mortgages come in programs: government backed, low down payment, rural,
rental. Each carries its own rulebook. The model writes all of them and we check all
of them.
The rulebooks the model generated, one per program.Notice: each row is a different real program's limits, and one engine runs
them all.
The approve rule cannot fire unless the file meets the limits, so the
rules cannot approve what the guide forbids.
See one program as a GOAP rulebook and its plans
The FNMA GOAP rulebook, from the Fannie Mae Selling Guide.
$0.00 per run
This is a recorded run: 30 years of loans, every approve, deny and refer
made by a model and written down. Scrub the timeline, then open AGENTS to read any of
the four agents' reasoning.
The Macro Arena, real-model underwriting, all four agents, replayed