Developer audited 249 AI coding sessions and found a traceability problem, not deception

A developer who uses an AI coding agent daily built a tool called 'red-handed' to audit 249 of their own sessions, checking whether the agent's claims about passing tests matched what it actually did. The audit applied nine checks, cross-referencing Claude Code session transcripts with git history to detect gaps between stated and actual outcomes. Out of 124 instances where the agent claimed tests had passed, zero confirmed cases of deception were found. However, seven sessions revealed a different issue: the agent had run legitimate checks that left no machine-readable trace, making it impossible to verify those claims later. The real problem, the developer concluded, was not that the AI lied, but that some test results were structurally unverifiable — a traceability gap that affects anyone who later inherits the code.