Three frozen audio encoders — MERT-95M, MERT-330M, MuQ-large — have never been told what a fugue is. They hear Bach's Well-Tempered Clavier the same way they hear his solo-violin partitas. This is what they heard. Click a sphere to listen.
Sixteen fugues — eleven WTC harpsichord, four solo-violin (BWV 1001 / 1003 / 1005), one passacaglia — sit in the tightest hull in the figure. k=5 NN purity 0.91, PHATE compactness 0.37. The encoder pulls fugues together across the violin/harpsichord timbre boundary. This is the strongest claim the project makes.
Op. 28 Preludes + Op. 10 + Op. 25 Études — 26 piano works whose dynamics envelope, rubato register, and voicing density form a Polish-Romantic hull with PHATE compactness 0.12, tighter than fugues. The encoder doesn't know Chopin, but it knows this isn't Bach.
k=5 NN purity 0.89 across 8 instrumentation classes. Audio encoders are brutally good at timbre. The wrinkle: BWV 1002's parallel violin / viola / cello transcriptions land as one tight triplet — same music, three timbres, encoder still pulls them together.
Bach's two WTC books form a single dense region. Every one of the eight closest cross-cycle pairs is a WTC I fugue paired with a WTC II fugue. Form outweighs cycle. WTC II Fuga I acts as an attractor — the closest out-of-book neighbor for five WTC I fugues.
Same k-NN purity, different geometry. MuQ's k-means picks up the fugue/non-fugue split as a direction in latent space; MERT only finds it locally. Switch encoders in the toolbar and watch the figure re-cluster — the structural finding is the same; the geometry is not.