20.5.1 A controlled benchmark proposal

20.5.1 A controlled benchmark proposal

A disciplined DIAL-MUSIC–0 experiment would withhold one reusable musical operation from a small declared theory. The corpus would include positive examples generated with the operation, negative examples that are merely unusual, and surface variants that preserve the operation under transposition, tempo, voicing, or instrumentation. The learner would have to:

1. detect a persistent quotient obstruction rather than high prediction error alone;

2. propose an operation equivalent to the withheld generator;

3. transport previously admitted motifs and constraints into the extended theory;

4. generate held-out works that use the operation coherently; and

5. pass independent evaluation by musicians as well as structural and perceptual tests.

Baselines should include a sequence model, a novelty-reward system, a style-confusion objective, and an oracle supplied with the missing operation. Success would not mean producing one pleasant tune. It would mean recovering a teachable musical operation that supports sustained creation. This is the musical analogue of inventing a paintbrush rather than producing a single new image.

Music also exposes DIAL’s limits unusually well. A grammar may certify that a piece is coherent without explaining why a performance moves a listener. Conversely, practices dismissed as violations by one theory may be central to another musical culture. The observer, doctrine, and admission community must therefore be recorded as part of the experiment rather than hidden behind a supposedly universal theory of music.