Build with Gemini XPRIZE · Submission 1034160

Franchise Edge — a franchise disclosure diligence engine

$500K, a personal guarantee, and hundreds of pages of legal disclosure on a 14-day clock. Gemini reads it and returns a full underwriting report in minutes — every figure citing the page it came from.

Start here

Demo — under 3 minutes
AI live in production, and the decisions it is and is not allowed to make.
youtube.com/watch?v=2qksyPsDQ9k
A complete report, free
The full paid output on a fictional filing — built by the same engine, not a mockup.
engine.foundersplinko.com/sample
Run your own filing
Upload any FDD. Extraction, underwriting and the cash ladder run end to end.
engine.foundersplinko.com
Source
Public repository, AGPL v3. Schema, financial layer, and the test suite that gates every commit.
github.com/jmwproductconsultingllc/engine-foundersplinko

The architecture, in one line

Gemini decides what the document says. Deterministic code decides what it means for your money.

What the model decides

What code decides

Two consequences. The same filing gives two buyers the same answer, auditable line by line. And a corrected formula reaches reports that were already sold, not only new ones.

Every figure is labelled by where it came from

Disclosed Derived Benchmark Inferred

Disclosed — stated in this filing. Derived — our calculation from disclosed figures. Benchmark — an industry range, shown because the filing does not disclose it. Inferred — an AI classification. The labels are enforced by tests, and the colours they render in are pinned by a test that reads the stylesheet, so an inferred figure cannot be made to look disclosed.

What the engine refuses to do

An FDD never discloses what a unit costs to operate. Where a figure is missing, the report says what is missing and where to look for it — never a zero, never an em-dash, never a plausible average from a comparable brand. Of the filings measured, five make no earnings claim at all; the report says so, tells the buyer the franchisor is legally barred from giving them numbers anywhere else, and hands them the Item 20 call list instead. It costs the sale. That is why the rest of the page can be trusted.

Measured, not asserted

Filings processed to a complete report83 of 83 · 100%
Runs failed because the model could not read the business0
Corpus measured22,619 pages · median 259 · longest 723
Automated tests gating every commit599 across 40 files
Production deployments in the period262 · mean build 28.9s
Price of a full report$199, one brand, no subscription
Arms-length revenue in the period$0.00 — stated, not implied

An operator checked it against the source document

“I own and operate two locations in a franchise system, and I've read plenty of FDDs. I ran one through Franchise Edge and then went back through the source document line by line to check it. It found the things that actually decide a deal: the full fee stack, the churn, who to call. Where I found something wrong, it was fixed and deployed within days. That's the part I'd tell another buyer about.”

Deven Patel · multi-unit franchise owner · arms-length user, not a paying customer

His audit is the most valuable feedback this product has received. It found the financial-condition section describing a profitable franchisor as running losses. That defect was then measured across every filing in the catalog, fixed in the producer and locked at render, and deployed inside the submission window.

The rule the engine is built on

A sentence that characterises a number must read the number.
A null is not a “no,” and it is not a “yes” either.
A failure that reports success is worse than a crash.

All three are enforced by tests rather than by review, because a rule enforced by human review does not survive contact with scale.