The way almost everyone picks retrieval components is to read each one’s reported gain off a benchmark table and assume the gains add. Across 450 configurations over eight benchmarks, with corpora from 5,000 to 200,000 documents, three techniques whose isolated measurements predict a 46.6% ranking improvement deliver 17.2% together — a 63% shortfall that grows to 72% once all five individually-helpful techniques are stacked. Two further findings land harder: ranking quality and answer faithfulness answer to almost disjoint halves of the pipeline, and the configuration that wins on a small corpus is the one that falls apart on a large one.