A new open source Python toolkit called Reverify pairs AI models with deterministic verification tools to stop them from inventing details about binaries and source code, according to its GitHub repository.
The project says models are confident but frequently wrong when reverse engineering binaries, fabricating struct sizes, offsets and function behavior. In tests across 71 Windows system files, a model's standard, unverified answer was wrong 97% of the time, according to the repository.
Reverify runs on a propose-and-verify loop: a model proposes a claim, and deterministic tools, including binary parsers for PE, ELF and Mach-O formats, disassemblers for x86, x64, ARM and ARM64, a CPU emulator and cross-reference analysis, check it against the actual bytes. Every claim gets a verdict of verified, refuted, inconclusive or observed, paired with the evidence used to reach it, the repository says.
The toolkit ships as an MCP server that exposes its verification tools to Claude Code, Cursor and other MCP-compatible agents, plus a command line interface and a Python API, according to the repository. It is released under the MIT license.
Reverse engineering is a fitting proving ground for a problem every agent-heavy workflow runs into: a model that sounds certain is not the same as a model that checked. Tooling that forces a verdict against real evidence, rather than trusting the model's own confidence, is the more durable fix than a bigger context window.