How it works

Two tools. One makes a record tamper-evident. The other records at the seam.

Two tools. The first makes a record impossible to change quietly. The second makes sure the record gets written at all, without depending on the agent to write it.

Ledger · free, open source

Other tools show you what your agent did. Ledger lets you prove it.

Each entry is locked to the one before it. If anyone edits, deletes, or reorders a line later, the record breaks and points at the exact line that changed. You own the file. You can hand it to an auditor and they can check it themselves without taking your word for anything.

pip install arcaeon-ledger

from arcaeon_ledger import Ledger
log = Ledger("agent.log.jsonl")
log.append({"tool": "payment", "amount": "49.00"})
log.verify()   # ok=True only if EVERY row was checked; ok=False names the
               # tampered line; ok=None means the scan was bounded (empty
               # file, or unchained rows skipped) - falsy, and not a green

python -m arcaeon_ledger.selftest   # check our guarantees on your own machine

There are five things it does not prove.

Adapter · free, open source

Your AI keeps its own notes. Adapter is the recorded line.

A dispatcher types notes into an incident. Those notes are a record, and they are usually accurate. But they are written by the person, about the call, and they contain what that person decided was worth typing. If something gets left out, nothing in the notes shows a hole. It just looks like a complete incident.

Then there is the recorded line. It captures the call itself, whether or not anybody types anything.

Both are records. Only one of them is independent of the person being reviewed. And when a complaint lands six months later, nobody pulls the notes. They pull the tape.

Your AI keeps its own notes. Adapter is the recorded line. It sits between the AI and the tools it uses, in its own operating-system process the AI does not control, and it captures every call the AI makes whether the AI writes it down or not.

We pointed it at our own logging tool. That tool's own notes said it did nothing that run. The recording caught five things it did. Not because our tool was lying, but because it was not writing down everything it did, which is the ordinary condition of every AI logging tool we know of, including the ones we build.

pip install arcaeon-adapter

python -m arcaeon_adapter --ledger seam.log.jsonl -- <your mcp server command>

Here is what it does not do.