Reverify

About Reverify

Reverify is a verification framework designed to ground AI outputs in factual evidence, primarily focused on binary reverse engineering and source code analysis. It operates on a "propose-and-verify" loop where a language model suggests claims (such as struct fields, offsets, or function behaviors) and a deterministic toolkit validates them against the actual artifact. The tool returns verdicts of `VERIFIED`, `REFUTED`, or `INCONCLUSIVE` based on evidence from disassembly, pattern matching, or execution in an emulator. ### Key Features - **Deterministic Verification**: Uses a pure-Python core for parsing PE/ELF/Mach-O, disassembly (x86/x64/ARM/ARM64), and CPU emulation. It can optionally upgrade to engines like Capstone, Unicorn, LIEF, Z3, and angr. - **Context Management**: Includes a `rollover` feature to prevent context drift in long AI sessions by handing off state to a file rather than relying on lossy auto-summarization. Supported environments include Claude Code, Codex, Gemini CLI, and OpenCode. - **Grounded Metrics**: Measures the information density of verified claims to ensure the AI is providing meaningful insights rather than trivial facts. - **Negative Memory**: Maintains a ledger of refutations so that fresh AI contexts do not repeat previously debunked hallucinations. ### Interfaces Reverify is available as a command-line interface (CLI) and as an MCP server for integration with agents like Claude Code and Cursor.

Summarised from the official page linked above.