cygnux-labs/pcm-protocol is the Proof-Carrying Messages protocol that attaches evidence, checkable offline by the recipient, to claims made by AI agents. It is not a new version or an official release, but a protocol proposal published in the repository. It is written in Python and follows the MIT license. Rather than accepting inter-agent messages on trust alone, it proposes checking the records that correspond to those claims as well.
The protocol records an agent's tool calls in an append-only Merkle log and places the inclusion proof for those records in the message. A Merkle inclusion proof is the material that confirms a record is present in the log. The recipient inspects the signature, the proof, and the record offline. According to the README, individual claims are classified as substantiated, unsubstantiated, or failed, while a message as a whole is classified as accepted, accepted with unsubstantiated claims, or rejected. The tracekit repository from the same organization is presented as a reference implementation.
The README reports that it covered 712 false claims out of 926 pre-registered message evaluations, and that only 10 of them were accepted. The rest were all classified as stale evidence. What can therefore be confirmed from this result is the verdict reached in that particular evaluation; it does not mean performance in a broader environment. The protocol's grounds for its decisions are concrete, but it is hard to broaden the scope of applicability based on this evaluation alone.