GitHub radar
Proof that an AI's answer really came from the model it claims
FlashProve is an Anthropic project that gives AI providers a way to prove a reply came from the exact model they promised, without revealing the model's secret weights.
It's a tool that lets an AI provider prove a given answer came from the exact model it promised, without exposing its secret weights (the model's internal settings).
When it helps
Useful if you want to understand how AI providers could prove they aren't quietly swapping in a weaker model. There's no ready-made tool here — it's an unmaintained research prototype.
Pros
- Rigorous proof: formally verified with Lean 4, a proof checker
- Checking a proof needs just an ordinary computer with bash and g++
- Tested on a real large model — Meta's Llama 4 Maverick
- Source code is open under the Apache 2.0 license
Cons
- It's a research prototype: not maintained, no contributions accepted
- Building new proofs yourself needs a data-center-grade NVIDIA H100 GPU
- There's no ready service or model you can actually use yet
How to set it up — step by step
- 1Open your coding agent (e.g. Claude Code) in your project folder
- 2Send the agent the text from the block below
- 3Wait while the agent builds the verifier (verifier/build.sh) — no GPU needed for this
- 4Check that the agent ran the example proof in examples/accepted-proof and got "YES"
- 5Ask the agent to explain in plain words what that example actually proves
- 6Don't ask the agent to build new proofs — that needs an expensive H100 GPU
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Open the anthropics/FlashProve repository, read the README and docs/what-is-proved.md, build only the verifier (the verifier/build.sh script — the lightweight part, no GPU needed) and check the stored example proof in examples/accepted-proof. Explain to me in plain words what it actually proves. Don't try to generate new proofs — that part needs an expensive H100 GPU.
Checking a finished proof just needs an ordinary Linux machine with bash and a g++ compiler, under 2 GB of RAM. Building proofs yourself (being the provider side) needs an NVIDIA H100 80GB, hundreds of gigabytes of RAM and disk, and hours of time — that is data-center scale, not something a home computer can do.
Open on GitHub▌ More finds