reverify: MCP server + CLI grounds AI claims in deterministic checks

reverify: MCP server + CLI grounds AI claims in deterministic checks

reverify went up on GitHub today: an MCP server and CLI that takes what an AI agent claims about a binary and checks it against the actual bytes, marking each claim VERIFIED or REFUTED with the evidence attached.

Reverse engineering is a bad place for a confident guess. A wrong answer about what a function does reads exactly like a right one, there is usually no test to run, and the person asking is asking because they can't tell. What makes the split work here is that ground truth is sitting in the file, a disassembler can settle "does this call recv" without a model's opinion. That is also the limit of the pattern. Propose-then-verify only helps where something deterministic can actually decide, and most of the domains people want it for don't have that.

The claim I want to see more of is the one about grounded facts surviving a context reset: verified findings kept outside the conversation, so a compaction doesn't quietly turn a checked fact back into a guess.

I'd read the REFUTED list first. The claims that got knocked down tell you more about the model you're pointing at the binary than the ones that survived.

https://github.com/2akouwu/reverify

#dev