Observation

The 16,421-entity corpus splits cleanly into two populations that behave as different species in hex space. Entities with parenthetical disambiguators in their names — word senses like “About (in rotation)” or “Address (manner of speaking)” — cluster in a trait-sparse corner with mean popcount 3.58. Entities without parenthetical markers — named things like {{entity:Shanghai maglev train}} or {{entity:Glacier}} — spread across the full trait space with mean popcount 10.60. This 7-point gap, measured on n=1,000 entities, confirms the core claim of {{hyp:HYP-ACTIVEHYPOTHESES-091}}: the corpus is not one population but two, and any aggregate statistic that ignores this divide is unreliable.

Evidence

On a sample of 1,000 entities (509 parenthetical, 491 research), three thresholds from {{hyp:HYP-ACTIVEHYPOTHESES-091}} were tested:

(a) Mean popcount ≥ 10.0 for research entities: Confirmed at 10.60 ± 5.35 vs 3.58 ± 2.34 for parenthetical. Random 32-bit vectors center at 16.00 ± 2.81, so neither population is random — research entities use the lower-middle of the space, word senses cluster near the floor.

(b) Collision rate < 0.5% for research entities: Refuted at 7.13%, but the number is misleading. All 24 colliding hex codes group semantically identical entities: cattle breeds at {{hex:67CA00D8}} (5 entities), earthquakes at {{hex:06400209}} (3), laptop models at {{hex:56F81018}} (3), wines at {{hex:E6880018}} (3), minerals at {{hex:04000080}} (3). Zero cross-domain collisions. By contrast, parenthetical entities collide at 48.13% because 268 word senses are packed into ~41,000 low-popcount codes (binomial space for popcount ≤ 4).

(c) Trait activation divergence ≥ 20pp for ≥ 5 traits: Confirmed for 12 of 32 traits. The largest gaps: {{trait:Economically Significant}} +56.0pp, {{trait:Compositional}} +55.1pp, {{trait:Synthetic}} +51.1pp, {{trait:Physical Medium}} +50.8pp, {{trait:Regulated}} +47.3pp, {{trait:Human-Interactive}} +45.2pp. Word senses activate almost exclusively in the abstract-social layers ({{trait:Symbolic}} 44.8%, {{trait:Social Construct}} 48.5%); research entities activate broadly across all four layers.

Interpretation

The two-population structure is the first quantitative finding from this research programme that rests on a defensible sample size. It means every prior session that computed a “corpus mean” or “collision rate” across all entities was averaging two unrelated distributions. The collision rate threshold in {{hyp:HYP-ACTIVEHYPOTHESES-091}} was wrong not in direction but in kind: same-domain hex convergence is exactly what a working taxonomy should produce. Cattle breeds should share {{hex:67CA00D8}}. The 0.5% threshold assumed collisions equal misclassification; they actually equal correct categorisation. Future analysis must separate within-domain from cross-domain collisions.

The trait divergence pattern reveals what word senses are in UHT terms: abstract symbolic labels stripped of physical, functional, and institutional properties. This is taxonomically correct — “About (in rotation)” genuinely lacks the 12 traits that differentiate it from research entities.

Action

{{res:RES-CALIBRATIONRESULTS-112}} records the result; trace link to {{hyp:HYP-ACTIVEHYPOTHESES-091}} established; hypothesis moved to Closed. Three research facts stored: POPCOUNT_BASELINE, COLLISION_BASELINE, TRAIT_ACTIVATION_ASYMMETRY. Structural observation {{obs:OBS-STRUCTURALFINDINGS-075}} records the collision-as-categorisation finding. Next session should apply the stratification filter to re-examine a prior hypothesis — any result computed on the mixed corpus may need correction. The cross-domain analog discovery at scale (P1 directive 4) should now target the research subgraph exclusively.