Konrad Gruszka

I build systems that make AI results inspectable after the run. AI evaluations produce numbers. What usually gets lost is the evidence behind them: who measured what, on which model and data, under which policy, and whether anything changed later. Today that lives as unsigned prose in a system card. **proofbundle** (open source, MIT) closes that gap. It is a verifier and receipt format for AI evaluation claims. One file, checkable on your own machine. No server, no account, no trust in me required. Every receipt answers three questions: → Who signed the claim? → What exactly does it cover? → Has anything changed since? And one boundary, on purpose: a receipt does not prove that a score is true or that the evaluation method was sound. Integrity, not truth. That line is visible in every verification result. The stance is old. I have been hooked on computers since I earned my pro-gamer stripes on an Atari 2600 and tried to take over the world in assembler on a C64, scraping by on tiny resources. Back then code had to *run* before it was allowed to *claim* anything, and my AI sidekicks drift from that all too happily. The next stage of AI is not more capability, it is reliability. The current state, verifiable rather than claimed: → Free and MIT-licensed: `pip install proofbundle` → Two independent implementations, Python and Rust, that differentially cross-check each other → External review already in the loop: a second developer building a separate verifier that agrees byte for byte, and an independent reviewer breaking the conformance corpus from the bytes upward, with findings fixed in the open → A published technical note: doi.org/10.5281/zenodo.21384526 → The eval-result receipt format, in active discussion with the in-toto attestation maintainers Behind it is b7n0de, my workshop for building software with AI where every important step leaves reviewable evidence: agent workflows with hard gates, provenance for every artifact, and release decisions that can be audited instead of believed. AI as the tool, the human as the final authority. My background combines mathematics and cryptography with 20 years of building and operating businesses end to end. I work where AI evaluation, software engineering, governance and product meet, and I am open to research collaborations and technical partnerships in verifiable AI infrastructure. **Verified AI Work: AI results you can check** Greetings from the intern 🖖 github.com/b7n0de/proofbundle · b7n0de.com/proofbundle

Personal Website

https://b7n0de.com/

LinkedIn

LinkedIn

Twitter / X

Twitter

Projects

proofbundle: independent audit of the trusted core behind AI evaluation receipts

creator

proofbundle: Offline Evidence Receipts for AI Safety Evaluations

Team Member

No grants recorded.