What this is. The whole session was one investigation wearing many hats \u2014 attention as distance, dimension, curvature, flow, topology. Every view came from a single eigendecomposition of one attention matrix. This instrument computes that decomposition once and reports all the facets together, on real BERT attention, with each structural claim auto-audited against its shuffle null. That audit is the point: it caught two overclaims I made earlier by hand (the curvature story, "the flow found the grammar"), and now it catches them automatically \u2014 the report says ARTIFACT when a number doesn't beat its null.
Read the report honestly. For BERT head 0: intrinsic dimension ~5, view coherence 72% (mostly visible in 3D), cluster persistence structured (fragments gradually), flow classification sorting tokens into pools/creeks/ocean. But curvature reads ARTIFACT \u2014 10% but below the ~19% shuffle null, so it is noise, not hyperbolic hierarchy. Head 1 differs where it counts: higher dimension (~8), lower coherence (50%), a diffuse persistence cliff. The real, null-surviving distinctions are dimension and persistence character; curvature is not real for either at this scale.
The one object. The 3D is the relational embedding every earlier instrument was a facet of \u2014 tokens placed by attention distance, colored by flow type, sized by depth. Toggle heads to watch the whole structure change; paste any N\u00d7N attention matrix to get its audited report. Green tier: real attention, exact decomposition, and \u2014 the actual deliverable \u2014 a report that runs its own null controls so it cannot flatter the data. That discipline, not any single view, is what the session was really building.