agentic-label-engineering (ALE)
I built a public CLI in which only `ale verify`, after running a task's acceptance commands itself, can mark an agent's work done. It rejected work in 13 of 26 real runs, but under my /work process it took 53 minutes and $8.92 (notional) where plain Opus took 7 minutes and $2.46 for the same hidden-test result, so I retired it.

