Music Analysis

An agent answered a question about Infest with 2,326 tokens of memory in its final context. Getting to that answer required thirteen model calls and 19,831 input and output tokens. The verifier accepted it. One of its claims attributed evidence from the fourth song to the first. That execution contains the central problem of this study. A smaller context can conceal a more expensive decision process, and an accepted answer can conceal an unsupported interpretation. Neither becomes visible if the experiment ends at the final prompt or the aggregate score. ...

An album ends. An agent has read every lyric. We ask it whether the ending transforms something established near the beginning. The agent can produce a convincing paragraph with almost no memory. That is precisely the problem. Fluency gives us very little evidence that it remembered the right songs, preserved their differences, or resisted turning a collection of voices into one convenient story. In my previous post on MLflow, I argued that a prompt belongs in the release process because changing an instruction changes application behavior. Here I want to bring that argument into my research on musical narratives. Memory selection changes behavior too. It needs an evaluation contract. ...

Abstract This study traces the lyrical evolution of the Beatles across four albums — Rubber Soul (1965), Revolver (1966), Sgt. Pepper’s Lonely Hearts Club Band (1967) and Abbey Road (1969) — using a pipeline of LLM-driven topic modeling (BERTopic), a vector database (OpenAI text-embedding-3-large indexed in ChromaDB), embedding geometry, and graph theory. The study puts a falsifiable question to the canonical critical narrative — that the Beatles moved through four discrete stylistic eras — by testing whether the albums occupy separable territories in lyrical-semantic space. ...

A song can change language while continuing an image, a joke, or an argument. It can also stay in one language while changing subject. Aquamosh, by Plastilina Mosh, offers a setting for asking how a similarity-based representation handles those possibilities. The question is narrower than whether a model understands the album: how do thresholded embedding similarities vary with language transitions, and where do they disagree with another model’s reading? What changed in this revision The previous title claimed a falsification of the distributional hypothesis. Its interpretation also treated GPT-4o-mini as equivalent to a human reference and generalized the results to commercial recommendation, moderation, and support systems. Those claims exceed the evidence available from this experiment and are withdrawn. ...

An album can revisit an idea through different words, voices, and images. It can also repeat the same words while changing what those words mean. That makes musical narrative an interesting setting for a measurement question: what does similarity between lyric embeddings actually tell us? The original analysis compared lyrics from Abbey Road and The Dark Side of the Moon. I expected a representation of sustained themes to reflect my reading of Pink Floyd’s album. That expectation is an interpretation to test, not a ground truth against which every different model result must be declared a failure. ...