Results

Hypothesis scorecard

Species-abundance distribution — H0 confirmed

The Preston-octave histogram is the textbook "hollow curve": 64% of species appear in exactly one playlist, 19.7% in two, 8.1% in three to seven. The popular-track backbone is bounded — the most-placed track appears in 365 playlists (~4.6% saturation) and the next 9 tracks are between 277 and 365, after which the SAD transitions cleanly to the log-series singleton-rich regime.

The empirical dominant signal: most tracks are rare, a small core is ubiquitous, and the curve fits a Hubbell log-series to 1.4 M σ decisiveness over a power law. The popular core (top ~1% of tracks) is the "niche" element — biologically, the top tracks in this dataset play the role of the "core-satellite" species documented by Hanski (1982).

Species-area relationship — H1 confirmed

The slope ratio 1.28 is above the dual-scaling prediction threshold 1.10. Both slopes are far above classical ecology's range [0.20, 0.35] — this is the island SAR signature, expected for the high-β-diversity playlist-habitat mosaic.

Taste-tribe communities — H2 REFUTED

Greedy-modularity (Louvain) community detection on the 47K-node track co-occurrence graph (3.4M edges, thresholded at co-occurrence ≥ 3) identifies 292 communities at Q = 0.586 — a strong modularity signal. The top 10 community sizes rank from 14,534 tracks down to 1,065; the long tail contains many small niche clusters of 30–500 tracks.

But here is the important finding: every one of the top 8 clusters has a within-cluster retention rate BELOW the degree-preserving null.

The null $\bar r(\text{null}_c)$ is the retention rate we'd expect if we picked 15 random degree-matched clusters of the same size. Because the heavy-tailed degree distribution puts popular tracks in every cluster at above-baseline probability, the null retention is high (0.91–0.98). The real Louvain clusters fall below that ceiling, more so for smaller clusters (which end up overindexed on rare-track tails).

This is the opposite of what a niche-confirming experiment would show. Three things are true at once:

1. The Louvain partition has structure (Q = 0.586, 292 communities, clear size hierarchy).
2. The clusters are NOT above-null-dense in playlist-cohort density.
3. The popular-track backbone dominates the null retention — any plausible assignment of tracks into clusters will look informative because the heavy-tail degree distribution rewards you with high retention by default.

The cleanest read: the listening-taste ecosystem is neutral-dominated under heavy-tailed degree and the Louvain clusters are a useful visualization, not a strong niche assertion.

4.5 Audio-features-conditional robustness (H2 stays REFUTED)

To verify the H2 refutation isn't a noise artifact of mixing audio and non-audio tracks in the same graph, we joined Spotify audio-features (maharshipandya/spotify-tracks-dataset, 89,741 unique tracks with audio) onto the MPD by track_id, then binned tracks by mean pairwise audio cosine similarity and re-ran the Louvain + retention test within each audio-similarity band.

H2 stays REFUTED — and strengthens with higher audio similarity:

The only bins where any cluster has η > 1.0× are the two bottom audio-similarity bins where the null baseline itself sits near 1.0 (0.98-0.99) — refutation is null-baseline dominated at low similarity and gets stronger as we condition on actual audio similarity.

Interpretation: the H2 refutation is not a noise floor; it survives the most rigorous "control for track-level similarity" we could put on this dataset. The refutation strengthens with audio similarity, which is exactly the opposite of what a niche-confirming experiment would show. Listening-taste co-occurrence clusters are heavy-tail artifacts more than they are deep carrying-capacity niches.

Synthesis

The listening-taste ecosystem is decisively mixed: neutral at the species-abundance level, niche-structured at the community-partition level (Q = 0.586), but anti-niche at the within-cluster density level (H2 REFUTED, enrichment 0.6–0.9×).

The classical ecological reading is that this metacommunity is neutral-dominated under heavy-tailed degree — the fact that community-detection algorithms partition the graph does not mean those partitions are ecologically strong. They are informative visualizations, but their per-cluster density deficit vs null is a warning: do not over-interpret Louvain communities on heavy-tailed co-occurrence graphs as deep biological-style niches.

Where the prior breaks (commitment-to-disconfirm)

Two pre-registered priors were broken by the data — both in ways that improve the picture rather than weaken it:

1. The power-law SAD prior. Initial expectation from music recommendation literature is that track popularity follows a Zipf- Mandelbrot power law. REFUTED: the SAD is decisively log-series, not power-law, at ΔAIC = 1.35 × 10⁶. The power-law appearance of the top ~1% of tracks in rank-abundance plots is decoupled from the bulk SAD shape; the heavy tail is a superficial feature of niche, not a fundamental property of the ecosystem.

2. The taste-tribe enrichment prior. Initial expectation (from classical niche ecology and from the Louvain Q = 0.586 number) was that clusters should exceed the degree-preserving null. REFUTED: clusters are 0.6–0.9× null. The dominant force is the heavy-tailed popularity distribution; popular tracks appear in many cluster-cohort playlists by virtue of their ubiquity, not their nicheness.