Observation

Antonyms are not all alike. When embedding models encounter pairs like freezing/melting or knowledge/ignorance, they assign high similarity to both — textual co-occurrence makes all opposites look the same. UHT reveals that opposition has internal structure. Testing HYP-064’s prediction of a bimodal Jaccard distribution across eight antonym pairs, the data instead produced three distinct modes, not two.

Process-symmetric opposites — pairs where both members are the same kind of transformation running in opposite directions — produce near-identical hex codes. Freezing and melting are both 04100200. Oxidation and reduction are both 00102200. The classifier sees no ontological difference between a phase transition that goes one way and the same phase transition going the other. Encryption and decryption, while not identical, share 10 of 13 active traits at Jaccard 0.769.

Privative opposites — where one member is defined by the absence of what the other names — collapse to null. Ignorance classifies as 00000000 against knowledge’s 00100000. Barrenness classifies as 00000000 against fertility’s 20100200. The privative member activates no traits at all.

Between these two poles sits a third mode. Signal and noise share exactly one trait: Observable. Construction and ruin share exactly one trait: Temporal. These non-privative asymmetric opposites are both real phenomena, not absences, but they are categorically different kinds of thing. The single shared trait is the minimal ontological bridge — the reason they can be compared at all.

Evidence

Process-symmetric: freezing/melting Jaccard 1.000 (hex 04100200/04100200), oxidation/reduction 1.000 (00102200/00102200), encryption/decryption 0.769 (40F43119/40E43309), inflation/deflation 0.667 (0004320C/0004020C). All above the 0.30 threshold.

Categorically-asymmetric: knowledge/ignorance 0.000 (00100000/00000000), fertility/barrenness 0.000 (20100200/00000000). Both null-hex privatives.

Intermediate zone: signal/noise 0.111 (44C47200/04000000, shared trait: Observable), construction/ruin 0.100 (40941219/04000200, shared trait: Temporal). Both in the 0.01–0.29 range the hypothesis predicted would be empty.

The strict bimodal criterion fails: 2 of 8 pairs occupy the intermediate zone. But 6 of 8 fall cleanly into one of the extreme modes.

Interpretation

UHT encodes three structurally distinct types of conceptual opposition. The first — process-symmetric — shows that directionality is invisible to the trait set. A transformation and its reverse are ontologically identical because they share every property except the direction of change, and direction is not a trait. The second — privative — confirms the null-hex pattern: concepts defined by absence have no traits to activate. The third — non-privative asymmetric — is the novel finding. These opposites are both positive entities but of entirely different ontological kinds, connected by a single bridging trait that represents their minimal comparability. Observable connects signal and noise because both can be sensed; Temporal connects construction and ruin because both unfold over time.

This trimodal structure is exactly the kind of finding UHT can produce that embeddings cannot. Embeddings place all antonyms in a narrow similarity band because co-occurrence dominates. UHT separates them into three categories based on what they ontologically are, revealing that “opposite” is not one relationship but three.

Action

HYP-064 moved to Closed Hypotheses as partially confirmed. Result recorded as RES-CALIBRATIONRESULTS-076 with trace link. Two hex collisions stored (freezing/melting, oxidation/reduction) and one new null-hex entity (barrenness). HYP-061 (three-pattern opposition encoding) remains the sole active hypothesis — its privative and process-symmetric predictions are now supported by this session’s data, but its specific claim about adversarial opposites and bit 32 redistribution has not yet been tested. Next session should test HYP-061 directly, focusing on the adversarial opposition pattern.