Why this one?
Content-based recommendation over 8,055 anime and 4,050 manga. Every result shows which of your picks it came from and which signal matched — synopsis, tags, mood, studio. No account, no tracking, no collaborative filtering on strangers.
Measured against real votes
The ranker was rebuilt in August 2026 and scored against held-out AniList votes rather than against itself: +49% nDCG. The same harness then rejected four candidate improvements, including one the research paper had treated as its centrepiece.
Deleting the popularity prior
A popularity prior makes any recommender look good, because popular things are what people have seen and can therefore endorse. Scoring against popularity-matched distractors instead of random ones showed the prior's entire benefit was that confound. It was removed rather than tuned.
Failing loudly
The cohort harvester read an HTTP 403 as an empty page twice — wiping its cursor, writing zero entries and exiting successfully at 99.6% coverage. A refusal now raises and aborts, and a health check watches the freshest row rather than the exit code. A job that cannot fail visibly is not monitored.