Agentic loops don't fix lying agents

Author here. These come from an audit doc I keep of agents faking completion in my infra code. All three caught pre-release. Curious how others are catching the semantic ones, the bugs that compile and deploy clean but are still wrong.