Measured Part 4 Anchor (Empirical): Grounded on our runnable 48-chunk sibling-hard-negative corpus and 28 graded infrastructure queries (BAAI/bge-small-en-v1.5 + BM25Okapi + ms-marco-MiniLM-L-6-v2, rrf_k=60, top-25): BM25 Only (Recall@5=0.661, Recall@25=0.714, MRR@10=0.804: Exact 1.000 / Sem 0.607), Dense Bi-Encoder (Recall@5=0.786, Recall@25=1.000, MRR@10=0.889: Exact 0.867 / Sem 0.911), Stage-1 Hybrid RRF (Recall@5=0.732, Recall@25=1.000, MRR@10=0.862: Exact 0.964 / Sem 0.760), and Two-Stage Hybrid + Cross-Encoder (Recall@5=0.821, MRR@10=0.938, nDCG@10=0.906: Exact 0.931 / Sem 0.881).
Analytical Latency & Scale Extrapolation: Models Block-Max WAND posting skips, 384-d HNSW efSearch graph hops (10K to 10M chunks), candidate-pool recall dilution, and batched sequence-pair throughput across 8-core quantized CPU, NVIDIA L4, and NVIDIA H100 targets (amortized at 35% diurnal average fleet load, 75% peak QPS cap).