How do you adjudicate between a research team wanting a...

The research proposal for a closed-loop system risks uncontrolled drift, while the data team’s 300-query set lacks the statistical power to resolve performance issues. The solution is to break the cycle by using a frozen external generator conditioned on the corpus and applying consistency filtering. The 300 real queries should be reserved as a calibration anchor that is never used for training. You must define a reversal condition: if the anchor’s overlap statistic drifts by more than a few standard errors between rounds, the loop is closing incorrectly and must be stopped.