Your reranker gets replaced with a strictly more accura...

The optimal context cap (k⋆) is derived from the width of the model’s safe zones (primacy and recency), which is independent of the reranker’s accuracy. Even with a better reranker, marginal candidates will still be demoted into the trough if k is too large. Since the new reranker has changed the rank-probability distribution, the previous cliff-based heuristic is invalid. The correct approach is to empirically re-measure the system’s performance to determine the new optimal k, rather than relying on assumptions about the reranker’s trustworthiness.