ProblemMost approval designs gate only the final irreversible action, so a Bot can spend an entire run building on a bad source and a human sees the mistake at the last step, when the work is already wasted.
TechniquePut three gates on a serious workflow rather than one. The source gate asks whether the inputs are the ones agreed. The evidence gate asks whether each conclusion is actually supported by what was gathered. The action gate is the familiar approval before anything irreversible. The first two can be checks the Bot performs and reports; only the third needs a human.
Why it worksIt moves detection earlier. A wrong source caught at gate one costs one step; caught at gate three it costs the run, and it trains you to approve tired work rather than reject it.
When to useAny workflow long enough that a bad input is expensive to discover late: research, outreach, reporting, anything that ends in a send.
When not toA short reversible task, where three gates is more ceremony than the work.