Every concept so far was real music theory. What if you inject a rule that doesn’t exist — a made-up law — and see whether the model can learn it at all? The answer draws a sharp line between a habit and a rule.
The invented law
Inject a tritone over the tonic — a colour that isn’t part of any style — into the corpus, and train. Then test two versions of the same made-up idea: a uniform habit (do it everywhere) versus a conditional rule (do it over this chord, something else over that one, nothing over a third).
Classical, with the invented tritone injected
The uniform version — a tritone over every major chord — was learned cleanly. The same mapping made conditional on which chord came through three times weaker, and one branch not at all.
Tritone-over-I rate: control 0.007; uniform-dose model 0.064 (z=3.6); conditional model at a matched 1.82% dose 0.023 (z=2.7 vs uniform). The V→blue branch: 0.005 vs a control 0.013 — not learned.
The verdict
A small transformer absorbs what note tends to appear far more easily than what note appears when. It prefers a statistic to a rule. This is a real limit on “teaching” — and it lines up with the metal result that a uniform subgenre habit transfers, while structure is harder.
Researcher notes
- Dose-matched. The conditional model was trained at the same 1.82% injection rate as the uniform one, so the three-fold gap isn’t a difference in exposure — it’s a difference in learnability.
- Framing. The tritone-over-tonic isn’t truly alien (it’s a Lydian colour); the point isn’t novelty but whether a chord-conditional mapping is learnable at small scale. It barely is.