Self-initiated synthetic product work. I built a deterministic testbed that checks whether an AI agent's approved tool, target, arguments, expiry, and one-use authorization still match the final validated tool call.

The suite covers target and argument substitution, context and ownership drift, approval expiry, replay across restart, and post-write readback mismatch. It blocks before tool entry and produces machine-readable evidence and regression reports.

No customer system, production credential, or vulnerability claim is involved. The public repository is a proof of method for the 48-hour Agent Tool Boundary Audit.