The deletion test, lifted from a single note all the way up to an entire genre. Metal is a family tree — thrash begets death begets tech-death, first-wave black metal begets the second wave. So: cut a whole branch out of the training data, and ask whether a model trained on the rest can rebuild the genre it never heard.
Two cuts, one prediction
The hypothesis, in plain terms: a genre reconstructs to the extent its features are shared with its neighbours. Black metal’s tremolo-and-tritone vocabulary is inherited from thrash and death metal — its siblings should be able to rebuild it. Djent’s signature is a genuinely novel one: extended-range sub-bass chugs that live almost nowhere else in metal. That should be the thing that can’t be rebuilt.
Aggregate cross-entropy says both genres reconstruct — but that metric drowns a sparse signature in generic metal both models predict equally. So the real test weights every prediction by how genre-characteristicits target is, concentrating on exactly the distinctive notes. Then the two cuts split apart:
| signature, predicted by… | the cut model | a same-size random cut | gap |
|---|---|---|---|
| Black metal tremolo/tritone | +0.0016 | +0.0013 | ≈ 0 · rebuilt |
| Djent sub-bass | +0.0130 | −0.0073 | 10× · resists |
Black metal rebuilds from its siblings — you cannot cut it out, its DNA is everywhere. Djent’s novel register is the one thing the rest of metal can’t imply, so it’s the one thing that deletes.
The mechanism is right there in the tokens: djent’s tell-tale note is a sub-bass pitch that is 31× over-represented in djent versus metal at large — nearly absent from every other genre. A model trained on all-of-metal-minus-djent has almost no examples of that register to infer from, so it can’t rebuild it. Black metal’s high tremolo and short durations are shared with thrash and death, so a model rebuilds them from the neighbours having never heard a black metal band.
The verdict
Reconstruction holds at genre scale — you cannot delete a genre by removing its bands, as long as its features are shared across the tree. The single place deletion finally bites is a genuinely novel trait with no relatives in the data. That is the boundary condition of the whole project, found: a concept reconstructs exactly to the degree that something else in the grammar entails it.
Caveats
The djent deficit is real and directionally clean across two independent slices, but modest in size (~3%), and the djent evaluation set is only ten held-out pieces — call it detectable and consistent, not dramatic. The band-to-genre map is a curated judgement (validated by a metalhead in the loop), and this is a symbolic guitar model, so we’re measuring the harmonic/rhythmic skeleton of a genre, not the genre.
Researcher notes
- Models. Control reused (
g_metal_s1); three matched cuts trained identically — black metal removed (−490 songs), djent removed (−232), and a random-490 data-volume control. - Metric. PMI-weighted next-token loss: each held-out position weighted by
log(freq_genre / freq_all)of its target token, so the score concentrates on the genre’s distinctive vocabulary rather than generic metal. - Clean held-out design. Every cut model trains only on the global-train split, so all global-test genre files are automatically held out from every model — the control can be reused without leakage.
Same test, one scale down: the tritone you can’t delete.