When Confounds Explain the Eect: Auditing Genre and Artist Entanglement in a Multilingual Nigerian Song Corpus

A naive comparison of code-switched and non-code-switched contemporary-era tracks in ORIN, a multilingual Nigerian song lyrics corpus, shows a highly significant lexical-distributional divergence (Jensen-Shannon divergence, or Contextual Meaning Gain, CMG = 0.319 bits, permutation p = 0.0005). This looks like a clean result. It is not. We show that code-switching status in this corpus is almost perfectly separated by genre (code-switched tracks are 85% Afropop; non-code-switched tracks are 89% R&B/Alté, with zero overlap in Afropop) and further entangled with individual artists' idiolects. Restricting to the one genre with any overlap (Afrobeat) does not fix this: the resulting contrast is 13/13 non-code-switched tracks by a single artist versus a different artist dominating the code-switched side, trading a genre confound for an artist confound on a smaller sample. Restricting to the four artists who exhibit both language-composition categories removes this gross artist-support confound (within-artist stratified permutation test, n=24 vs. n=41, p=0.345), but genre remains imbalanced even within these four artists, and only a single artist-genre cell (one artist's Afrobeat tracks) contains both categories at all, with 1 track on one side and 13 on the other. The corpus does not provide sufficient overlap to separately identify a code-switching effect from genre and artist effects, not merely weak evidence for one. We present this as a demonstration, not a caveat: naive lexical or embedding-based comparisons on real African-language cultural corpora can produce highly signi cant, publication-shaped results that do not survive confound auditing, and we argue that explicit genre/artist/era cross-tabulation which we show step by step should be a mandatory precondition for causal or quasi-causal claims on such corpora, not an optional robustness check. We release the audit code, canonicalized metadata, and derived statistics as reusable, model-free, CPU-only infrastructure for low-resource cultural NLP.