A review pipeline posts every finding directly to pull requests. The team wants to send uncertain findings to a human reviewer instead, without losing real issues. Which design fits best?

Answer: D — Add a verification pass reporting confidence per finding, thresholds tuned on past outcomes

A verification pass that self-reports confidence alongside each finding enables routing, and checking that confidence against real outcomes sets the threshold. Discarding low-confidence findings (a) loses real issues instead of routing them. The generating session (b) is anchored by its own reasoning. Severity alone (c) ignores a usable signal once it has been calibrated.