The obvious design stores the prompt, and then the audit trail becomes a second copy of every piece of customer data the system ever saw — with the retention and disclosure obligations that follow. A hash proves the input was what someone later claims it was without holding it, which is the whole point of an audit record and none of the liability. Recording the model version alongside is the other necessary field: a decision is not reproducible without knowing which model made it, and model versions move underneath you. One tool, doing one thing that most systems bolt on badly.
A single-tool audit ledger for AI decisions: it records the decision along with the model version and a hash of the inputs.
- record_decision writes the decision to the ledger with the model version that made it and hashed inputs rather than the inputs themselves.
A ledger for it to write to.
One command plus a key — npx -y audit-ledger-mcp, then supply credentials