Bitcoin AI Provenance: signed receipts for what an agent did, anchored where nobody can edit them

Bottleneck

When an autonomous agent acts on your behalf, four questions come up the moment something goes wrong. What was asserted? Who signed it? What did it depend on? Was that committed somewhere nobody can quietly edit? Most agent frameworks answer none of them.

System

BTC AI Provenance Core answers all four with a dependency-free reference implementation: Ed25519 identities and domain-separated signed envelopes, deterministic canonical JSON with strict input rejection, SHA-256 content-addressed artifact storage, signed provenance receipts with cycle-checked lineage graphs, domain-separated Merkle batching with inclusion proofs for every receipt, and compact 37-byte commitment payloads anchored through a Bitcoin Core wallet.

Control

• Policy before action. Signed agent policies with session-key binding enforce model, tool, target, daily-spend and action-count limits.
• No replay. Strictly increasing nonces; exact satoshi accounting.
• Honest claims. The threat model states what Bitcoin does and does not prove. The local development journal is hash-chained and never presented as an anchor.

Result

An evidence layer that any Blocpod system can adopt when a client needs proof of agent behavior that outlives the vendor. It is public so that the claim “the system records what it did” can be inspected rather than believed.