Tamper-proof verification infrastructure for AI evals and scientific claims
Tamper-proof verification infrastructure for AI evals and scientific claims
Project Details
Updated 07/07/26 · Provided via application · VerifiedHow do we know if a claim, whether scientific or an AI evaluation, can be trusted?
Who tests the claim?
Is the test independent?
ValiChord seeks to answer these questions, with each claim independently tested blind, and the testers unable to change their results afterwards. Trust, with ValiChord, is architectural and not just policy.
When a researcher makes a claim, their result is cryptographically sealed, and only their data and methodology is sent to validators, whether human or AI. Only when all the validators have also sealed their results are the actual claim and the validators' results revealed — a process called commit-reveal, so no one can get a last-move advantage. The results are published in a tamper-proof form that I call a "Harmony Record".
The grant will allow me to continue working on ValiChord for the for the next 12 months: improving and stress-testing the commit-reveal protocol, running validation rounds, and continuing to foster relationships in the AI safety and AI governance fields.
I have already made a demo of the decentralised commit-reveal at work. It can be seen here — https://www.linkedin.com/feed/update/urn:li:activity:7466596558680018944/ — and you can use it yourself here: https://valichord-demo.onrender.com/demo
At the moment, I am the sole developer of ValiChord.
ValiChord represents a major career pivot for me. I am a trumpet teacher by trade, and this is the second pivot I have made recently, as I also dipped my toe into documentary filmmaking. My documentary of Amelia Earhart's first transatlantic crossing has won several awards at film festivals across the globe. You can watch it here: https://www.youtube.com/watch?v=C3MwFDJbytc
I am dyslexic and used AI for spelling and punctuation corrections for this answer; the content and words are my own. AI has been a godsend to me as it has ensured that my communications are now more likely to be taken seriously, which has not been the case for much of my life.
The project repository can be found here - https://github.com/ValiChord/ValiChord
Theory of Impact
Updated 07/17/26 · By grantmaking.aiDecisions about whether an AI system is safe to deploy increasingly rest on evaluation results. Right now, those results are self-reported. The world takes the lab's word that the test was run properly, that nothing was cherry-picked, and that awkward results weren't quietly left out. This is the same structural weakness that produced the scientific replication crisis — except here the failure mode isn't a retracted paper, it's deploying a dangerous system on the strength of a safety test that was massaged after the fact.
The problem compounds between rivals. Labs and governments cannot coordinate on safety — slow down, share commitments, verify treaties — if none of them can check each other's claims. Every framework for international AI governance quietly assumes verification that doesn't currently exist.
ValiChord reduces x-risk by making eval integrity structural rather than procedural. When results go through commit-reveal, validators test blind, verdicts are sealed before anyone reveals, and the record cannot be edited afterwards — not by the lab, not by the validators, not by me. Gaming a benchmark, selectively reporting, or tidying up disagreement stops being discouraged and becomes impossible. That gives mutually distrusting parties — labs, auditors, states — something they currently lack: a way to agree on what a model actually demonstrated without having to trust whoever ran the test.
People
Updated 07/17/26 · By grantmaking.aiTeam Member
Discussion
No comments yet. Be the first to share your thoughts.