Reverify
A tool that checks every structural or behavioral claim an AI makes against ground truth with deterministic tools, so a claim about a binary, an API, or a piece of code only counts once it is verified rather than asserted.
Reverify makes a deterministic tool the judge of an AI's claims: the model proposes a claim, such as what a struct's fields are or what a function does, and a tool checks it against the actual artifact, returning verified or refuted with evidence, rather than letting the model assert a fact on its own. It also addresses context rot on long tasks: instead of a lossy automatic summary, a rollover command hands a session off to a file and starts a fresh one, so a long task does not drift and does not need to be manually cleared.
The hardest place to prove this works is binary reverse engineering, where hallucination is worst, so that is where the project's benchmark numbers come from: on 71 real Windows system files, an AI's unaided textbook answer was wrong 97% of the time, while Reverify caught every wrong claim and never accepted an incorrect one, a result reproduced independently on aarch64 and re-run in continuous integration on Linux and macOS on every push.
Its deterministic core covers PE, ELF, and Mach-O parsing, x86, x64, ARM, and ARM64 disassembly, pattern scanning, CPU emulation, and protobuf and TLV parsing, exposed as an MCP server and a CLI for Claude Code, Codex, Gemini CLI, and OpenCode. It is written in Python and MIT licensed.