PatchProof: Coding agent patches you can prove
AI coding agents are impressive right up until they aren't: costs balloon, patches break things, and nobody can say why. PatchProof makes every patch verified, reproducible, and auditable, whether it came from a live LLM or a curated candidate.
- Type
- Internal R&D
- Field
- AI Agents · Program Repair
- Services
- AI & Data Analytics
- Release
- Published
The challenge
Left alone, a coding agent will burn tokens on dead ends and declare victory on a patch that fails half the test suite. Real codebases need guarantees the raw models don't give: hard verification, bounded cost, and a trail showing how each fix was found.
Our approach
Every candidate patch runs through hard verification gates in an isolated workspace, from scope checks and secret scanning to visible tests and hidden regressions. Failures come back as bounded evidence that guides the next attempt, and every run records its full configuration, so results are reproducible and models can be compared fairly.
What this delivered
- Gate-verified repair loop for every candidate patch
- Bounded evidence packets that guide each repair
- Live LLM generation under the same controller
- Reproducible benchmarks with caching and budgets
- Model-matrix evaluation across providers
- Full provenance, traces, and audit artifacts