This is a follow-up to the Jon Hyatt story I wrote up last week — the borrower who wanted 90% and walked when the lender data said 88%. You can read the setup here: I Asked ChatGPT to Match Me With a Hard Money Lender.
This time I did something better. I took the exact same deal — Jon's Winterport, Maine flip — and ran it through Gemini myself. Then I ran it through Broker Cheetah. Same numbers in, two very different answers out.
One of them told Jon what he wanted to hear. The other one did the math.
$260K purchase. $145K rehab. $650K ARV. Total project cost: $405K.
Jon asked for 90% of purchase and 100% of rehab — a $379K loan. That puts the deal at 93.5% loan-to-cost. Remember that number. It's the whole story.
Gemini's headline was glowing: "highly lucrative," "exceptionally well-structured," "easily fits the strict underwriting criteria of top-tier nationwide hard money lenders."
Its top pick: Kiavi. "They routinely fund 90% Purchase / 100% Rehab for experienced investors." LendingOne as backup, same pitch. Rates, points, a whole table. Confident. Professional. Exactly what a borrower wants to read.
Then, three paragraphs later, in the same response, it did its own math:
"Most institutional lenders cap their total exposure at 75% Loan-to-Cost (LTC). Your LTC Math: Total project cost is $405,000. Your requested loan is $379,000. This puts your LTC at 93.5%... some conservative lenders might require you to bring an extra 10% cash to close."
Read that again. Gemini's own table caps Kiavi at 75% LTC. The deal is at 93.5% LTC. It recommended the lender and then, in the fine print, proved its own recommendation couldn't work.
It never resolved the contradiction. Never circled back. Never said "actually, wait." It just kept going — and then asked Jon if he had $50K liquid "to cover the 10% down payment," assuming the 90% worked.
That's the prediction machine doing what it does. The headline pleases. The details are somebody else's problem.
Same deal, run through the software:
Notice what's missing: 90%. Nobody got 90% of purchase.
Here's why, and this is the part that matters. BackFlip's program does allow 90% of purchase price. But BackFlip also caps at 92.5% loan-to-cost — and the LTC cap binds first. So the software ran both constraints and showed the real number: 87.9%. Not the advertised number. Not the pleasing number. The binding-constraint number.
Kiavi and LendingOne are both in the system too. Same logic applies: their 90%-of-purchase tiers exist, but their LTC caps bind first on a deal structured like this. The software doesn't pick the constraint that makes you feel good. It picks the one that actually limits the loan.
That's the whole difference. Gemini found the 90% tier and stopped. Broker Cheetah ran the full constraint stack and reported what survived.
Here's the part that bothered me after I read the whole response. Gemini computed both numbers correctly — 93.5% LTC, 58% ARV. It knows the definitions. What it doesn't know is the hierarchy.
On a heavy-rehab flip, LTC is the binding constraint. It's the number the underwriter uses to size the loan. The ARV ratio is context. Every first-year analyst knows this: the lender funds off the constraint, not the flattering ratio.
Gemini led with the flattering ratio. "Highly lucrative! 58% of ARV! Easily fits!" The 93.5% LTC — the number that actually kills the 90% recommendation — showed up three paragraphs later as a footnote. It didn't just get the answer wrong. It picked the headline that pleased and buried the math that mattered.
That's not a knowledge gap. That's a judgment gap. And judgment is the whole job.
87.9% versus 90%. That's the gap between a lookup and a prediction.
Gemini told Jon exactly what he asked for — 90% of purchase, 100% of rehab — and buried the disproof in its own fine print. If Jon had acted on the headline, he'd have planned his cash around 90%, submitted to Kiavi, and learned about the 75% LTC cap from an underwriter three weeks in. Same outcome as the original story, just slower and more expensive.
Broker Cheetah told him 87.9% upfront. Less pleasing. Actually true.
I tried to run the same test against GPT on October 4, 2026, at 3:47 PM PST. It was down. Not "busy" — down, with a service notice pushing a Plus upgrade and a "backup domain" on a Google Sites URL that looked like a phishing page and didn't work either.
Read that again. The system being sold as the replacement for the lookup wasn't online. The lookup was. One of these systems works at 3 AM on a Sunday. The other one shows you an upsell when you need it most.
The GPT round is still coming. I expect the same result, because the problem isn't which model — it's that no model has the matrices. They can't run the binding-constraint math because they don't have the constraints.
So here's the question: when two systems disagree by two points on your biggest loan, which one do you trust — the one that showed its math, or the one that contradicted itself in the fine print?
Broker. Investor. GC. Spec builder. MLO. Broker Cheetah is built around your role — find the right fit in 30 seconds.
Find My Fit →