Can a mood written in plain words find a show nobody has rated yet?

A retrieval system for anime that ranks by mood instead of genre, and can place a title nobody has watched yet. Viewers describe mood at length in reviews, so a frozen MiniLM encoder feeds two learned towers trained with InfoNCE on review segments against title metadata. The item tower never sees reviews, because a new title has none; titles are masked out of queries, so the model cannot match on the name; and each batch holds a title at most once, so a repeat is never pushed apart from itself.

The evaluation harness was written before the models, on purpose. Cold start is the headline split, and beyond-accuracy measures (coverage, Gini, serendipity and popularity lift) count as much as NDCG. In-catalog, a from-scratch BPR matrix factoriser with hand-derived gradients wins on accuracy and surfaces 2.9× more of the catalog. Two negative results are reported as plainly as the wins: the trajectory model the project set out to build was tested with a shuffle control, failed it, and was deliberately not trained.