agent-delivery-gates asks AI coding agents to prove their work

Field notes

The project treats an AI coding agent’s delivery report as a set of proof obligations rather than a narrative to accept at face value. Its checks include mutation testing, pre-change test execution, and deliberate removal of declared error handling. The homepage reports 103 dated gate-tally entries from September 7–9, 2026, and says the system uses a sandboxed dependency that cannot open a socket during gate runs. Seven worked examples link into the repository for inspection.