Legal requires that every retrieval decision be explain...

First, verify the evaluation set: if it consists of sentence-based questions, it likely ignores identifier queries where the dense-vs-BM25 margin is negligible. Second, separate the requirements: reproducibility is handled by versioning, while per-term attribution requires a fixed-basis (sparse) leg. The optimal design is a hybrid approach where the sparse leg provides the explainable score and the dense leg provides recall via rank fusion. If legal insists on dense-only, clarify that the auditability requirement is being dropped, not met.