Observation

UHT classification is not decomposable from an entity’s stated essential properties. When semantic-triangle extracts what an entity IS — its core properties, category, ontological status — and map-properties attempts to predict which of the 32 traits those properties would activate, the prediction captures only 20% of the traits the classifier actually assigns. The aggregate F1 across 10 entities from 8 domains is 0.275. The classifier activates roughly twice as many traits as property decomposition predicts (average popcount 9.5 vs predicted 4.3), with an average Hamming distance of 10 bits between predicted and actual hex codes.

Evidence

Ten entities tested: {{entity:Photosynthesis}} ({{hex:34753209}} actual vs {{hex:12400001}} predicted, 10 false negatives), {{entity:Candle}} ({{hex:C6D80008}} vs {{hex:00400000}}, 8 FN), {{entity:Democracy}} ({{hex:00BFFEFF}} vs {{hex:00002811}}, 18 FN), {{entity:grief}} ({{hex:00108281}} vs {{hex:8A200382}}), {{entity:wedding}} ({{hex:0080DAAA}} vs {{hex:000000FA}}), {{entity:Gold}} ({{hex:04008080}} vs {{hex:08000000}}, zero trait overlap), {{entity:supply chain}} ({{hex:40A43299}} vs {{hex:10048101}}), {{entity:Symphony Orchestra}} ({{hex:46FC7098}} vs {{hex:04000000}}, 14 FN), Gold (coins) ({{hex:C0800008}} vs {{hex:00080609}}), Deep Grief ({{hex:00101200}} vs {{hex:004086A9}}).

The three traits most invisible to property decomposition: {{trait:State-Transforming}} (missed in 6/10 entities), {{trait:Compositional}} (6/10), and {{trait:Intentionally Designed}} (6/10). These are relational and structural properties — they describe what an entity does within a system, not what it is made of or how it appears. The classifier infers them from holistic semantic understanding; simple property-to-trait keyword matching cannot reach them.

Democracy is the extreme case: 22 active traits, only 4 predicted. Properties like “participation of the citizenry” and “rule of law” trigger {{trait:Rule-governed}}, {{trait:Regulated}}, and {{trait:Ethically Significant}} — but miss 18 traits including {{trait:State-Transforming}}, {{trait:System-integrated}}, {{trait:Compositional}}, {{trait:Symbolic}}, and the entire social layer. The classifier understands that democracy is a self-modifying compositional system with signalling, institutional, and identity dimensions — none of which follow from its listed properties.

Interpretation

The classifier is not a property matcher. It performs 32 independent holistic judgments about what an entity IS, drawing on semantic understanding that exceeds what essential properties alone convey. The gap between map-properties prediction and actual classification quantifies what might be called the “holistic inference contribution” — the fraction of trait activation that arises from the classifier’s implicit world model rather than from surface-level property correspondence.

This makes the semantic-triangle plus map-properties pipeline a viable diagnostic tool: for any classified entity, the pipeline reveals which traits are “obvious” from stated properties (high map-properties confidence) and which are “inferred” (active in actual classification but absent from prediction). No embedding system offers an equivalent decomposition — embeddings are opaque vectors with no trait-level interpretability. This diagnostic is uniquely enabled by UHT’s structured trait architecture.

The dominance of relational traits in the missed set ({{trait:State-Transforming}}, {{trait:Compositional}}, {{trait:Intentionally Designed}}) suggests these traits carry the most implicit semantic load. They are the traits most dependent on the classifier understanding the entity’s role in a larger system, not just its intrinsic properties.

Action

{{hyp:HYP-CLOSEDHYPOTHESES-097}} refuted: recall 0.20, well below the 0.40 refutation threshold. {{res:RES-CALIBRATIONRESULTS-114}} recorded with full data. {{obs:OBS-STRUCTURALFINDINGS-077}} notes the diagnostic application is viable. Research fact stored identifying the three relational traits as a functional archetype of “holistic inference” traits. Next session should investigate whether the holistic inference contribution correlates with popcount or domain — do complex entities always have larger gaps, or do some simple entities also defy property decomposition?