PaperTrail

AgentLedger is an accountability layer for AI agents. When an AI agent acts on someone's behalf — reading documents, looking up records, sending messages, updating data — there is usually no reliable record of what it did, whether it was allowed to do it, or who approved it. AgentLedger makes every agent action pass through a safety pipeline before it can execute: (1) a policy check against role-based permissions and rules, (2) review by a second, independent AI model (the "guardian") for risky actions, (3) human approval for high-risk actions, and (4) execution only after all gates pass. Every step — proposals, blocks, guardian verdicts, approvals, executions — is written to a tamper-evident, hash-chained audit log (SHA-256, append-only SQLite), so any later edit to the log is detectable and the exact altered entry can be named. The demo shows three things: a normal work-order task completed with human approval; a prompt-injection attack hidden inside a document, stopped twice by two independent layers (policy, then the guardian model, which explains in plain language that instructions inside documents are not instructions from the user); and a tamper test where editing one audit record directly in the database is caught by the built-in "Verify audit log" check. Built solo during the hackathon; all demo data is synthetic.