I’m Richie Sater, a CPA with no math degree, and I recently used AI agents to solve a Soviet-era math problem that had been unsolved since 1986.
What’s interesting is that even after I solved it, convincing professional mathematicians to take the result seriously was a big hurdle.
They eventually trusted the result after I gave them what was essentially audit evidence: a test and transcripts they could verify themselves.
This experience made me realize: AI is producing more technical work than experts can possibly review, so every field is going to need a way to test that work at scale. That is what CPAs already do: we provide assurance over work we did not create and explain why someone else should trust it.
What happened
In 1986, Viktor Monakhov put “Problem 10.34” into the Kourovka Notebook. A couple weeks ago I sent him an email letting him know I’d solved it. After 40 years, the problem that had outlasted a generation of mathematicians finally had a proposed solution. But there is one issue: I don’t have a math degree. I have never taken a class in group theory. I didn’t know what the Kourovka Notebook was until a few weeks ago.
I built a system that runs AI agents from my MacBook, a small army of AI math workers. One searches old papers. Another generates examples. A third tries to break everything the others find. A fourth handles citations, and so on. It’s messy and noisy and dangerously effective.
After I sent Monakhov the first preprint, he and Irina Sokhor picked another problem and sent it over. A challenge. I copied the question into my system. The terminal lit up. My Mac mini started heating up as the agents divided the work.
At 11:18 that morning, I wrote back to Monakhov and Sokhor: I had a counterexample that answered their new problem. I attached a short note and hit send.
“We have significant doubts.”
How could a problem so important be solved in 12 hours by someone who can’t do group theory? They were suspicious, and rightfully so. This is where the accountant part of my brain kicked in.
In audit work, if someone hands you a number that looks too good, you don’t just nod. You trace it back. You rebuild the calculation. You design a test that will fail if the claim is wrong. I did exactly that: created an audit system to make it very easy for others to test the work.
“The GAP file together with the output transcripts have convinced us.”
Why I think this matters to CPAs
AI speeds up the production of raw output (In this case mathematical claims), but a thousand polished wrong answers can bury a single good one under a mountain of review work. My process forces the agents to build an explicit audit trail: verifying every citation, generating standalone executable code, and running adversarial checks before a human ever sees a line of text. It hands the reviewer verified evidence, not slop.
I think CPAs are going to be surprised at how good they are at managing agents. But I also think that CPAs are going to be the ones who figure out the assurance layer for a world where output is no longer the constraint. AI can produce an enormous amount of work. The constraint becomes expert review: experts suddenly have a tremendous amount of output to check.
That is the part that feels very familiar to me as a CPA. We already have a profession built around deciding what evidence is enough to trust work we did not create, testing it, and communicating that conclusion to other people. The bigger story isn’t really about math. Math is just the example I happened to live through.
Where it stands
I have posted a public preprint claiming a negative answer to Kourovka Notebook Problem 10.34. The underlying paper and verification materials are public and submitted for publication..
Preprint: arXiv:2608.02970
Verification materials: github.com/RichieSater/kourovka-10-34
A related first-person piece is now in the Mathematical Association of America’s Math Values editorial process. The editor wrote that the piece is a “strong fit for Math Values” and that she would “love to get this published.” We are currently working on edits and she has offered me the September 10th publication date, which I am accepting.
On September 30, I’m bringing the entire story – and the system architecture behind it – to NC State’s Algebra & Combinatorics Seminar.
Sign in to get access to this free resource, and all of our whitepapers and reports.
Download this content today!
Register Now Already registered? Click here to Log In
