Observation

The cleanest separation UHT has produced across 77 sessions: agentive concepts and non-agentive concepts occupy entirely non-overlapping regions of trait space. Not a single non-agentive entity activates as many traits as the least-activated agentive entity. The gap between the groups — 8.67 bits on average — dwarfs the 2-3 bit differences found in previous calibrations testing markedness, state-vs-process, or valence. Agency appears to be the ontological property UHT encodes most decisively.

Evidence

Six agentive concepts — predator (10 bits, 27730200), teacher (13, 050C4AF9), parliament (14, 4084FAD5), virus (11, A6321211), entrepreneur (9, 0100D2A9), immune system (12, 23775200) — averaged 11.5 active bits. Six non-agentive concepts — granite (5, 8E000008), wavelength (1, 04000000), entropy (2, 00002200), latitude (0, 00000000), alloy (6, C6001008), viscosity (3, 04003000) — averaged 2.83 bits.

The minimum agentive count (entrepreneur, 9) exceeds the maximum non-agentive count (alloy, 6). Direction holds in all six paired comparisons. Within-group agentive Jaccard averaged 0.263 versus cross-group Jaccard of 0.067 — a delta of 0.196, nearly double the 0.10 threshold. Latitude joins the null-hex population, consistent with the pattern that pure measurement coordinates lack any activatable ontological trait.

Interpretation

UHT appears to encode agency through cumulative trait activation: entities that can initiate action, respond to stimuli, or pursue goals activate traits across multiple ontological dimensions simultaneously — they are compositional, rule-governed, output-producing, state-transforming. Non-agentive entities, being passive substances or scalar properties, activate at most one or two structural traits (typically physical-substrate markers like those seen in granite and alloy).

This finding subsumes and explains several prior results. The markedness effect (session 71, +2.875 bits for marked antonyms) was partially an agency effect — marked concepts like war and pollution involve active processes. The state-vs-process discrimination (session 74, +2.1 bits) was a weaker version of the same signal, since processes inherently involve more agency than states. The null-hex pattern for measurement quantities (wavelength, latitude, voltage, pressure) reflects the extreme non-agentive end of this spectrum.

The zero overlap between groups is the strongest claim: if UHT consistently places all agentive concepts above 8 active bits and all non-agentive concepts below 7, this binary threshold could serve as a practical agency classifier — something distributional embeddings cannot provide, since agency is not a surface-level co-occurrence feature.

Action

HYP-ACTIVEHYPOTHESES-039 created and confirmed, linked to RES-CALIBRATIONRESULTS-038. Hypothesis moved to Closed. Research fact stored as FUNCTIONAL_ARCHETYPE “agency-encoding.” Latitude stored as NULL_HEX.

The next session should test whether this boundary holds under adversarial conditions: non-agentive concepts with superficially active names (e.g., “active ingredient,” “driving force,” “leading indicator”) and agentive concepts with passive framing (e.g., “dormant volcano,” “idle process,” “latent virus”). If the boundary survives linguistic framing variations, agency encoding is a structural property of UHT, not a lexical artifact.