b7n0de/proofbundle

proofbundle is a small Python tool that creates a tamper-proof receipt for an AI evaluation result. When someone reports a safety score or a capability benchmark for an AI model, there is usually no way to check the claim yourself, you just have to trust the lab that published it. proofbundle turns that kind of result into a single signed file that anyone can check offline, without a server or network connection, to confirm two specific things: that a stated party actually signed those exact numbers, and that nobody has quietly changed them since. It is important to understand what this tool does not do. A receipt does not prove the number itself is true, that the evaluation was designed well, that the issuer was honest, or that the underlying computation was correct. It only proves authorship and integrity, similar to how a cash register receipt proves a transaction happened and was not altered, not that the store treated the customer fairly. The README is upfront that this narrow guarantee is the entire point of the project, not a shortcoming. The technical approach relies on standard cryptographic signatures (Ed25519) and a tree-based structure called a Merkle tree that lets someone verify a large set of results piece by piece without needing everything at once. Optional features include selective disclosure, so a threshold can be proven while the exact score stays hidden, transparency log support so signed claims can be tracked publicly over time, and a per-sample auditing feature that lets an outside auditor spot-check random samples from a run. It also supports pre-registering an evaluation plan before running it, to make cherry-picked results after the fact visible. This is aimed at people building or reviewing AI evaluation pipelines: researchers running benchmarks, auditors checking published safety claims, or teams integrating with existing evaluation frameworks like inspect_ai or pytest. It installs via pip and includes a one-command demo that shows both a valid receipt and a set of deliberately tampered ones failing verification.