What I built
Provenant is a reference implementation for AI-agent audit trails: tamper-evident, automatic, examiner-ready records that prove what an autonomous AI agent actually did. It's a local, air-gappable layer that sits between an agent and the real world and seals every action it takes — no SaaS, no telemetry, no account.
- Tamper-evident log — an append-only, hash-chained attestation record (HMAC-SHA256) where editing, reordering, deleting, or truncating any entry visibly breaks the chain.
- Runtime policy gates — deterministic rules that allow, block, or require human sign-off before a high-risk action runs, with every decision signed into the same chain.
- Examiner-ready reporting — compiles a single self-contained HTML report (integrity verdict, signed history, approvals, policy decisions, and a compliance framework crosswalk) that opens anywhere and prints to PDF.
- Independently verifiable — optional external timestamping (OpenTimestamps to Bitcoin) lets a third party prove when a record existed without trusting the vendor, the server, or the key.
- Automatic, not reconstructed — capture is wired in through hooks, so records are sealed as work happens rather than written up after the fact.
Built for the compliance conversation
Provenant maps its records to the frameworks that are starting to ask "prove what your AI did" — the EU AI Act (Article 12), ISO/IEC 42001, SOC 2, the NIST AI RMF, and Colorado's AI Act (SB205). The crosswalk is control-level pointers, not legal advice — but it turns a vague "it's all logged somewhere" into an examiner-ready story.
It's sovereign by design: standard-library only, runs offline from a USB stick, and the signing key never leaves the machine.
Open source coming soon
Provenant is being opened up as a reference implementation — open source coming soon. The goal is a credible, self-contained way for any team running AI agents to answer the auditability question, without buying a platform to do it.