Peer-Isolated Recommendation: When Content-Only Personalization Beats Collaborative Signal, and When It Doesn't

We present Essence, a personal embedding cluster framework for recommendation that operates exclusively within a single user's engagement history, without consulting cross-user interaction data. We evaluate Essence on three datasets spanning music, e-commerce, and film (Last.fm-1K, Amazon Books, MovieLens-25M) against nine baselines, including collaborative filtering, popularity, recency-based retrieval, and two hand-implemented multi-interest neural baselines (MIND, ComiRec), with $K$ selected per dataset on a validation split rather than fixed in advance ($K=10$ Last.fm-1K/Amazon Books, $K=15$ MovieLens-25M). Across these three datasets, Essence's relative standing is associated with observed collaborative-baseline strength; establishing this as predictive would require additional datasets and an independently-defined domain property. Recency-Weighted, a non-clustering baseline, shows the identical pattern, indicating this belongs to the broader architecture family rather than to $K$-means clustering specifically. Where the evaluated collaborative baselines perform weakly, Essence ranks 2nd of 10 on Last.fm-1K and 3rd on Amazon Books, significantly beating collaborative filtering, popularity, and both neural baselines, and not significantly different from most recency-only baselines; where they perform strongly (MovieLens-25M), it ranks 8th of 10, losing to the same systems by the largest effect sizes in this study. All primary baseline and stratification comparisons use paired bootstrap significance testing with FDR correction and BCa cross-checks. A targeted stratification test finds $K$-means clustering, isolated from recency-weighted retrieval, provides no measurable benefit even for the users it should help most: no stratum shows Essence significantly ahead of recency-weighted retrieval alone. Essence is a data point on the domain conditions under which peer-isolated, content-only personalization is or isn't the right architectural choice.