On the QASC validation split ($n = 926$), the strongest strict-parity retrieval system the paper reports is ColBERTv2/RePlug at R@$10 = 85.0\%$ (95% paired-bootstrap CI $[83.4, 86.6]$). Because ColBERTv2/RePlug adds an extra learned reranker on top of the matched dense interface, it is reported under the SP+ type rather than the pure strict-parity SP type, so its headline number is interpreted alongside, not directly against, the SP systems. Strict parity still holds the encoder, candidate pool, cutoff $k = 10$, matching rule, and split policy fixed across systems, so the comparison isolates the retriever + reranking pipeline.