Q15 · How do you detect and mitigate "lost in the middle"?

Detect by measuring it on your own eval set rather than citing the paper: hold the evidence set constant and force the gold chunk into position 1, the middle, and last. The spread is your position sensitivity. It varies by model and by task, so somebody else's U-curve is a hypothesis, not your number.

Mitigate in order of cost. Keep k small — fewer chunks in the middle at all, and it saves money too. Order by reranker score then interleave, putting the two highest-scoring chunks at the head and the tail. Restate the question briefly after the evidence, which puts the task in the strong end position for about fifteen tokens.