October 13, 2026: on 5,183 SciFact research documents and 100 sampled queries, search ranking quality rose from 71.76 to 76.23 points out of 100 when keyword and meaning-based results were combined, a 6.2% relative improvement. This score rewards relevant documents appearing near the top of the first ten results; it is not the percentage of correct answers. Combination weights were fitted on separate query folds. A separate 300-query comparison replaced the MiniLM encoder with a pinned BGE-small model using full precision: meaning-based search ranking quality increased from 64.51 to 70.05 points, an 8.6% relative improvement. An 8-bit version separately reached 67.76 points, a 5.0% gain. The encoder improvement did not improve the combined search score, so the existing combined ranking remains the baseline. Saved rankings were independently checked against unchanged relevance labels. The 300-query test had been inspected in earlier experiments; it was not an untouched test set. Timings overlap other work on a shared CPU machine.

Based on source review and saved project artifacts. Recorded measurements apply to the stated model, dataset and configuration; they are not claims of a current production deployment.