The default way we check AI output is to ask another model whether it looks right. That second model has the same blind spots as the first. I've built a working prototype of the alternative: a small trusted checker that verifies a claim by executing it, and returns a signed result anyone can re-run. Refuted means an executed counterexample. Confirmed comes tagged with exactly how much assurance it earned. Abstain is a real option, never a fake pass.
On a labelled set of smart-contract vaults, ten independent LLM auditors flagged a safe vault as vulnerable ten times out of ten, and missed a real bug 40% of the time. My checker got the same set with zero false positives and zero false negatives, and produces a witness you can re-run yourself. Wherever an output has checkable structure (code, math, structured data, rules, contracts), executing the check beats grading it with a bigger model.
Open-source the kernel, publish a benchmark across several domains, and work on the open problem that gates the whole idea: getting a usable specification cheaply for outputs that don't come with one. If it doesn't pan out, I'll publish that too.
Single-author prototype, not shipped, audited adversarially twice (only cosmetic overclaims found, no soundness holes). The hard part is unsolved and I'm saying so up front. I pre-register what I expect and retract when it fails: I dropped this program's earlier framing after five pre-registered tests came back negative.
Independent researcher based in Italy, working on AI you can verify rather than just trust. 2026 papers on representation limits and substrate scaling (on Zenodo). I also build AionNexus (industrial predictive maintenance) and NormaAI (EU compliance tooling). github.com/Dan23RR, daniel.culotta@gmail.com
someone arrives first
0