On a possible concrete mitigation strategy. I wrote up a pattern that seems to be working for me as I push through a brownfield migration. It's a little mental model file. Always human edit. Agent checks for falseness only. I started calling it `HITL.md` https://warrenwhipple.com/hitl https://github.com/warrenwhipple/hitl Curious if you've seen anything like it. Or whether you think it can scale to teams/PR workflows.