Notes You Can’t Delete
depth
findings/Interpretability
Looking inside

The structure we caught ourselves on

The model lays the twelve keys out in a circle of fifths. Tempting to call that discovery. We ran the control — it's the data.

data, not discovery

The model represents the twelve keys with circle-of-fifths geometry: keys a fifth apart look similar inside it, tritone-apart keys look opposite. It is enormously tempting to call this the model discovering music theory. This page is us catching ourselves.

Run the control

Before crediting the model with a discovery, ask whether the structure was already in its input. It was. The same circle-of-fifths geometry sits in the raw pitch-class statistics of the music — more cleanly than in the model — because keys a fifth apart simply share more notes. Anything that represents which notes are present inherits the circle of fifths for free.

fifth-apart / tritone-apart similarityin the raw notesin the model
a fifth apart (should be similar)+0.50+0.13
a tritone apart (should be opposite)−0.63−0.28
The geometry is stronger in the raw pitch-class statistics than in the model. Correlation between the model’s key layout and the raw one: 0.30.

Reported as what it is

Not a discovery — a property of the data the model faithfully carries. The rule we took from it, and applied everywhere after: before claiming a model found a structure, measure whether that structure was already in its input. The same discipline is what deflated the drum result into an honest “overdetermined” and forced the targeted genre probe.

Researcher notes

  • Baseline. Build a 12×12 key-similarity matrix from raw pitch-class histograms (transpose each piece to all twelve keys), and compare it to the model’s learned key-embedding similarity. The raw one shows a cleaner circle.
  • Why. Adjacent keys on the circle of fifths share six of seven scale tones; a tritone away they share only one. Any note-presence representation reproduces that automatically.