Three pre-registered hypotheses. Two confirmed by neutral log-series and dual-scaling SAR; one refuted by the audio-conditioning check.
prediction · z_small / z_large > 1.10
At small sample sizes, curves climb steeply because each new playlist is likely a fresh tribal sub-sample; the slope flattens as we draw more samples from the heavy-tailed backbone.
prediction · ≥ 5 clusters > 1.5× null
Within every top-cluster, real retention is below the degree-preserving null. Holds across all 5 audio-similarity bands — refutation strengthens with audio similarity.
prediction · ZSM AIC < logseries + 10
Log-series decisive at the abundance level. Community partition rejection (H2) shows: abundance ≠ community structure — listening ecosystem is mixed, not pure-neutral.
The corpus is 20,000 Spotify playlists drawn uniformly from the Million Playlist Dataset Challenge, holding 1,338,693 track-playlist incidences across 259,652 unique tracks. After filtering to tracks seen in ≥ 3 playlists and playlists containing ≥ 5 such tracks, the analytic dataset shrinks to 19,406 playlists × 65,029 tracks.
All paper numbers derive from this analytic subset unless otherwise noted. The figures page uses the same underlying track/playlist bipartite.
The MPD listening ecosystem occupies a mixed regime. Across the species-abundance distribution, neutral log-series fits decisively. Across the community partition, Louvain finds 292 taste-tribe clusters at Q = 0.586 — strong modularity. But across the within-cluster retention test, every top cluster sits below the degree-preserving null. The carry-capacities that biological niches promise are an artifact of the heavy-tailed popularity backbone; the clusters are useful navigation, not carrying-capacity islands.
We document H2's refutation honestly and pre-register the audio-features robustness check that strengthens it.
Four atmospheric covers (Grok Imagine, vision-verified clean for trademark exposure) and the four core figures from the paper.
