lin-0215
17.6 MovieLens as a conjunctive gate
Experiment: RADAR MovieLens conjunctive gate
The chronological MovieLens 100K study uses only training events to construct geometry and joins [ Harper and Konstan , 2015 ] . Apex RADAR exceeds matched rating-weighted UMAP by \(0.0272\) normalized discounted cumulative gain at rank ten (NDCG@10), reduces incidence in all five seeds, retains at least \(98.4\% \) of every anchor, and closes cocycles to numerical precision. Those structural and ranking gates pass.
Its rating root-mean-square error (RMSE) is \(1.154\), however, which is \(0.154\) worse than the 2D matrix-factorization comparator and fails the preregistered \(0.05\) safeguard. The correct outcome is diagnostic, not a positive recommendation claim.
A RADAR realization is promoted only when incidence strictly decreases, private bands and non-collapse anchors remain feasible, induced cocycles close, the apex signal survives, and every task-specific safeguard passes. Passing ranking while failing calibration blocks the aggregate claim.