The Quadrilingual Probe: How Aquamosh (1998) Falsifies the Distributional Hypothesis Across Five Embedding Architectures
Abstract This research uses Aquamosh (1998), the quadrilingual debut album by Plastilina Mosh (Spanish, English, French, Japanese; produced by Tom Rothrock and Rob Schnapf — Beck’s Odelay team), as an empirical falsification probe for distributional sentence embeddings. The album’s quadrilingual structure converts code-switching from anecdotal concern into a quantitative experiment: every language transition is a guaranteed lexical discontinuity, allowing us to dissociate topical continuity from surface form. Core Finding (CONFIRMED): In all five sentence-embedding architectures probed — OpenAI text-embedding-3-large (3072-dim, decoder), Google LaBSE (768-dim, encoder, parallel-corpus), BAAI BGE-M3 (1024-dim), multilingual-E5-large (1024-dim), and paraphrase-multilingual-MPNet (768-dim) — a language switch in consecutive lyric lines approximately doubles the probability of “window break” (the embedding similarity falling below a calibrated coherence threshold). Mean relative gap across models: 1.69×; range: 1.31× (E5) to 1.94× (OpenAI). Permutation tests against H₀ of language-rupture independence reject with z = +6.54 (OpenAI), z = +4.51 (LaBSE), both p < 10⁻⁴ over 10,000 simulations. Logistic regression with GEE clustered by track and controls for line position and anchor/successor languages yields OR = 3.99 [2.51, 6.36] for OpenAI (p < 0.001), OR = 2.52 [1.39, 4.57] for LaBSE (p = 0.002). LLM-as-judge against GPT-4o-mini shows OpenAI declares “rupture” while a sophisticated reader sees continuity 3.18× more often in switches than in same-language transitions (false-break rate 0.060 vs 0.191). ...