The earlier high-D test was synthetic — flagged as such. This one runs on a real corpus: 4,674 documents, 14,731 TF-IDF dimensions of actual language. The structure's cut is really two cuts: the eight subject verticals (finance·health·IT·law…) and the four rhetorical modes (ethos·pathos·logos·mathea). I tested each against real text — and they land differently.
Real documents route to their domain at 81.4% across 8 classes. Pairwise, distinct subjects (health vs commerce, gov vs commerce) hit 94–96%; even similar subjects (two computing topics) still separate at 91% — real vocabulary is rich enough that the threshold is cleared easily. The divergence gradient the synthetic sweep predicted is present, just with a high floor.
The mode proxies — numbers (mathea), questions (logos), emphasis (pathos), pronoun-stance (ethos) — are largely independent of subject: topic explains only 0.5–18% of their variance. So the 4×8 grid really is two orthogonal dimensions, as designed. But independence isn't usefulness — whether the modes carve outcomes needs mode-labelled text, which a topic corpus doesn't have. Untested, not disproven.
Two axes, two verdicts. The subjects sort real language cleanly — that half is done, and it agrees with what the whole industry already ships. The modes are real and independent, but a topic corpus can't tell you whether they predict; only mode-labelled text can. Nothing here says the structure fails — it says half of it is proven and half of it has one honest experiment left.
— ROOT0, with AVAN. subject cut passes on real text, mode cut awaits its labels.