I'm a Master Electrician. Three months ago I couldn't write a line of code. So this is not a mathematics paper, and I'm not the one who could referee it if it were. I handed a session a pure-math research program — entropic optimal transport, total positivity, embeddings into the positive Grassmannian — that I cannot check one line of. On purpose. The question was never whether an AI can do mathematics. It was whether the provenance discipline I've been building could keep the work honest in a domain where I have zero ability to verify it.
The discipline is the same idea the trade runs on: you don't trust the tradesman, you trust the inspection trail. Every artifact hashed the moment it's written. Every step appended to a trace that can't be rewritten. The proofs attacked by fresh sessions with no memory of how they were made. Every session closing by naming its own overstatements. And nothing called “validated” without a third-party witness. The mathematics is a test weight — something heavy enough to stress the process, not the point of the exercise.
It caught two real errors the same day it produced the work. One was a formula written down wrong — a threshold that the text claimed depended on temperature when the actual proof says it doesn't; it was masked because the first numbers happened to be run at the one value where wrong and right coincide, and a fresh audit at other values exposed it. The other was a genuine flaw in a proof, which an attacking session didn't just flag — it built a numerical counterexample, forced a repair, and left the original claim standing only as a labelled conjecture. Seven pieces were audited; six came back sound, the seventh drew blood. A six-of-seven record with one caught-and-repaired flaw is more believable than a clean sweep would have been.
The part I'm proudest of is the idea it killed — my own. Partway through, the most exciting thought of the day showed up: formalize the discipline itself into a theory. I ran that idea through the same audit I run everything through, and it came back “this already exists, under several other names.” So I parked it. A system that will dismantle its own author's most flattering idea for zero cost is the strongest evidence I have that it isn't just telling me what I want to hear.
Here's the ceiling, stated plainly, because leaving it out would defeat the entire point: no human mathematician has checked any of this. It's carefully-audited, fully-traceable candidate mathematics, waiting for a referee. Every number resolves to a hash, and the paper deposits its own sealed provenance trace, so you can verify the record for yourself. Either the mathematics survives human review, or the trace shows exactly where it broke. Either outcome is information.
→ The paper (preprint + sealed trace, CC BY) · the code · the methodology · the experience report · the self-audit · the Fable 5 case study