How do you adjudicate a debate between increasing top-k...

Increasing top-k improves recall, but it often places more documents in the model’s ‘trough’ of performance, which can hurt end-to-end accuracy. Do not guess which effect dominates; instead, measure the accuracy-by-position curve at k=20 directly. If the trough dominates, the solution is not to revert k, but to use post-retrieval techniques like truncation or reordering to ensure that high recall and positional safety are not mutually exclusive.