The problem PatchProof solves

AI can write code incredibly fast, but it can also confidently make a change that looks correct without actually being correct.

ProofPatch makes that process safer. Developers can ask an AI to fix a bug, add a feature, or make a security change, while ProofPatch independently checks what actually changed, makes sure protected tests were not tampered with, runs the tests, and records the evidence.

Instead of asking “Can I trust what the AI said?”, developers can ask “Can I verify what actually happened?”

For teams, this means safer AI-assisted development, easier code reviews, and a permanent cryptographic record of verified changes that can be anchored to Ethereum for later verification.

Challenges I ran into

The hardest part was making sure the AI couldn't simply change the tests to make itself look correct.

At first, the verification loop treated a newly created test file differently from an existing one. That created an edge case where an AI could potentially introduce or modify test code and then pass verification.

I had to rethink the baseline: what existed before the AI started making changes had to be captured first and protected throughout the entire run. I also added checks to make sure the actual files on disk matched what the AI originally proposed.

Another hurdle was making the system handle failures honestly. If the AI claimed that everything worked but the independent test run failed, the evidence had to win over the AI's claim.

I eventually built the verification flow around a simple rule:

The AI can propose. It cannot decide whether its own work is correct.

That became the foundation of PatchProof.