Three Solved Problems, Zero Proofs: The AI Math Story Crypto Briefing Didn't Verify

Crypto Briefing published a headline that should have required a forensic chain of custody. An AI system, the report claimed, solved three open mathematical problems. No model name. No paper link. No formal proof artifact. Just a claim moving through a Web3 media outlet with the gravitational pull of a verified result. Silence in the logs is louder than any statement. When a story with this much impact arrives without a single verifiable artifact, the correct response is not excitement. It is chain-of-custody inspection. I have spent the last decade auditing cryptographic claims in whitepapers and DeFi protocols, and this pattern is familiar: high-impact assertion, missing metadata, and an audience asked to trust the headline rather than the evidence. Let me be precise. FrontierMath, the benchmark designed by Epoch AI, was created to test AI on research-level mathematics. Public evaluations from its early days showed mainstream models struggling to solve even single-digit percentages of problems. These are not exam questions. They are problems designed to resist pattern matching and require genuine mathematical construction. A claim that any AI has crossed from benchmark failure to solving open problems — not answering contest questions, but resolving unsolved research questions — would be a step change in machine reasoning. That demands exceptional proof. What did Crypto Briefing provide? Three information points, all repeating the headline. No solved problem was named. No verification method was described. No independent mathematician or formal theorem prover was cited. The phrase “Open Problems benchmark” itself is ambiguous. It could be a separate benchmark. It could also be a subset of FrontierMath where the difficulty bar is unknown. If those 50 problems were curated as approachable open subproblems rather than the canonical millennium-class questions, the word “solved” loses its center of gravity. In my due diligence work, I treat unsupported claims the way I treat unaudited oracle price feeds: suspicious until the contract-level evidence appears. During the 2020 DeFi exploit investigations, I traced a $15 million loss to a flawed oracle integration hidden in EVM bytecode. The project’s public post-mortem pointed to a “liquidity imbalance.” The actual metadata — transaction logs, timestamp ordering, and price update calls — told a different story. The same discipline applies here. Before accepting that three open problems have been solved, I need to see the answer transcripts. I need to see whether the solutions are natural-language narratives or machine-checkable proofs in Lean, Coq, or Isabelle. I need to know the compute budget, the model parameters, and whether the training set already contained fragments of the target problems. That last point is not pedantic. Benchmark contamination is the quiet killer of AI claims. If the model’s training data included discussion of these open problems, the system may be retrieving and recombining known partial results rather than producing original proof paths. The claim of “solving” could also be a media compression of something more cautious: a promising construction, a conjecture-strengthening example, or a counterexample that reshapes a hypothesis without closing it. None of those are trivial. None of them are equivalent to a formal proof. The article, as parsed, did not tell us which one happened. There is also the missing denominator. The report says the AI solved three problems. It does not say how many problems were attempted. If the model failed on the other 47, the “success rate” of six percent is not a revolution. It is a prompt engineering curiosity. Publishing only the numerator is the oldest statistical trick in the book. I have seen the same move in token launches, where projects highlight one profitable trade while hiding the rest of the trading history. The image is static; the provenance is a phantom. Always ask for the full distribution. Now let me give the bulls their due, because my job is not to dismiss the direction, only to demand evidence. The underlying trajectory toward AI-assisted mathematical discovery is real. Even if Crypto Briefing’s story collapses under verification, the broader movement toward hybrid systems — large language models paired with symbolic computation and formal proof assistants — is gaining momentum. Theorem provers like Lean are moving from academic niche to industrial infrastructure. The toolchain that lets an AI propose a lemma and a verifier confirm it is already being built. That is the actual story hiding beneath this unaudited headline. The implications, if a verified instance ever emerges, are structural. Mathematical research itself would shift from proof production toward problem formulation. The bottleneck would become asking the right questions and designing testable formal specifications, not grinding through derivations. Automated theorem proof checking could become a standard layer in scientific publishing. Peer review would need a new protocol: one part human judgment, one part formal verification. And math education, which still relies on handwritten exams and homework sets, would face a visible evaluation crisis. An AI that can solve open problems will solve any take-home assignment ever written. That transformation will not wait for journalism to catch up. But none of that transforms this article into evidence. The article gives us no audit trail. No model weights. No proof files. No link to a verified repository. It gives us a claim in a headline and a vortex of speculation. In my experience auditing whitepapers, the projects that have real cryptographic breakthroughs do not publish press releases first. They publish proofs first. They let the mathematics speak before the marketing team does. A 2017 ICO I audited claimed homomorphic encryption in its consensus layer. Two weeks of reading showed the construction was mathematically impossible. The team retracted after my proof-of-concept repository circulated. The lesson stuck: real technical claims survive artifact inspection; fake ones hope you never ask for the artifact. So here is the test that matters. If an AI genuinely solved three open problems, the solutions exist as formal objects. They can be checked. They can be compiled. They can be reproduced. Until those artifacts are public, the headline is sentiment, not news. Metadata whispers what the contract screams. The contract here says: no proof, no provenance, no consequence. Demanding that evidence is not cynicism. It is the only professional response. Give me the Lean files. Give me the independent verification committee. Give me the failure rate on the other 47 problems. Then we can talk about what the future of mathematics looks like. Until then, this story is a checksum that does not match its source file. Re-run the hash, and let the silence in the logs tell you what the headline refuses to say.