Training-Free Binary Projection Filtering for Dense Retrieval: An Empirical Study of Candidate Reduction, Ranking Stability, and Failure Risk

Dense retrieval pipelines often rely on large candidate pools before reranking, making candidate generation and downstream scoring a practical bottleneck. This paper studies training-free binary projection filtering as a lightweight pre-filter for reducing the candidate set before dense reranking. Rather than presenting it as a universally superior retrieval method or a validated speedup technique, we ask a narrower practical question: how far can the candidate pool be reduced before average top-rank quality, retained relevance, and query-level reliability begin to break down? We evaluate the approach on five BEIR datasets: SciFact, NFCorpus, FiQA, ArguAna, and TREC-COVID. The revised evaluation compares exhaustive Dense retrieval, FAISS-HNSW, FAISS-IVF-Flat, and Binary+Dense retrieval, and includes projection-dimension ablations over Db∈{128,256,512,1000}, candidate-budget ablations over K∈{50,100,200,500}, five-seed robustness analysis, and typo-perturbed queries. In addition to MRR@10, nDCG@10, and Recall@100, we report filter-stage metrics including Retained@K, catastrophic failure rate, and Best Relevant Survival. Across datasets, Binary+Dense often remains close to exhaustive Dense retrieval in average top-rank metrics at representative operating points, but the filter-stage behavior is strongly collection-dependent. Larger Db and K generally improve retained relevance and reduce catastrophic failures, but they also increase filtering cost or reduce the degree of pruning. The latency results show that structural candidate reduction does not translate into consistent end-to-end wall-clock speedup in the current Python 3.16/NumPy implementation. Taken together, the results suggest that training-free binary projection filtering is best understood as a calibration-sensitive pre-filter and failure risk analysis mechanism rather than as a replacement for Dense or ANN retrieval.