Notes You Can’t Delete
depth
Deletion

The blue note you can't delete

Erase every blue note from the training data. The model still expects it, exactly where a soloist would bend one.

97–98% rebuilt

The blue note is the friendliest possible target for deletion. It’s not structural — it’s a stylistic inflection laid on top of the harmony, a bent third or seventh a soloist reaches for. If any concept could be erased by removing its instances, it’s this one. It reconstructs anyway.

What the detector hears

Before any deletion, the instrument has to be trusted. These are ear-check clips: the chord alone, then the flagged note held against it, rendered on clean sine tones on purpose — a semitone clash is far easier to judge without a piano sample smearing it.

Three flagged degrees, held against the chord

0:00 / 0:00

The ♭5 is why ♭5 was dropped from the definition: to the ear it reads as a lydian or altered tension, not a blue note. The human in the loop overruled the arithmetic.

Renderer: an additive sine synth in numpy, WAV via the standard library, ffmpeg to MP3 — no soundfont. The detector: half-second chroma windows, chord templates, a simultaneity gate so an arpeggio isn't mistaken for a chord.

Remove every one, and it comes back

Take all 456 jazz solos, raise every blue third and seventh out of them, and train a model that never sees one. Then, on held-out real solos, at the exact positions a soloist bent a note, measure how much probability the model puts on the blue note.

98.5%
blue third, rebuilt vs control
97.0%
blue seventh, rebuilt vs control
~0 bits
difference in expectation, sign flips across seeds

A model raised on jazz with every blue note edited away still expects the blue note, where a real soloist would place one, at essentially the control’s rate.

The verdict

We chose the blue note specifically because it seemed the most deletable — a decoration, not a structural necessity. It reconstructed anyway. The concept doesn’t live in its instances; it lives in the surrounding root, third, and key, which the edit never touched. This is the same result the metal tritone gives, in a completely different harmonic world.

Researcher notes

  • Ground truth. The Weimar Jazz Database ships chord symbols per beat, so the blue-note label is arithmetic on real chords, not a guess — the one judgement left to a human is whether a flagged note sounds blue.
  • Probe. Chord-relative likelihood on held-out solos, paired and piece-clustered. The 97–98% figure is the ablated model’s expected mass on the degree, over the control’s, at the positions the real solo used it.
  • Depth. The blues model was d256/6L/ctx512 with 12-key transposition augmentation on the training split only.