Observation

Hex-code proximity in UHT creates neighborhoods that span 2.5 times more domains than embedding proximity. When you look up what’s “near” Cathedral in embedding space, you find Seville Cathedral, Notre-Dame, Milan Cathedral — the same kind of thing in the same domain. When you look up what’s near Cathedral in hamming space, you find Cultural Astronomy, Language of Science, Polonia — entirely different domains that share Cathedral’s ontological property profile. This pattern holds across 15 entities from ecology, biology, music, architecture, political science, computer science, geology, physics, electronics, religion, genetics, and food science.

Evidence

Twenty entities were sampled across 8+ domains. For each, the top 5 neighbors by hamming distance and by embedding similarity were retrieved. Domain diversity was counted as distinct domains represented in each neighborhood.

Mean hamming neighborhood diversity: 4.33 domains. Mean embedding neighborhood diversity: 1.73 domains. Ratio: 2.50 (confirmation threshold was > 2.0). Seventy-three percent of embedding neighborhoods were single-domain — Digestive Enzyme finds only other organs, Acer Predator finds only other Acer products, Democracy finds only other instances of itself. Only 13% of hamming neighborhoods were single-domain.

Two exceptions broke the pattern. Supply chain showed diversity of 5 in both metrics — its embedding neighbors included nervous system, judicial system, and orchestra, suggesting the concept is so structurally generic that even distributional proximity captures its cross-domain nature. Base (transistor base) showed higher embedding diversity (5) than hamming diversity (3), driven by polysemy: embedding space found baseball base, anatomical base, nucleotide base — disambiguation artifacts, not genuine cross-domain discovery.

Five entities returned empty results in one metric (neuron, Leaves, Pottery, Metaphor, High Fermentation Beer had no hamming or no embedding neighbors), likely due to sparse neighborhoods in the respective metric spaces.

Interpretation

This is the clearest quantitative evidence yet that UHT hex codes encode something orthogonal to distributional semantics. Embeddings answer “what appears near this concept in text?” — which is almost always more of the same domain. Hamming proximity answers “what shares this concept’s ontological property signature?” — which systematically crosses domain boundaries. The 2.5x ratio is not an artifact of noise; the hamming neighbors often make structural sense (Digestive Enzyme near Human Emotion and Short-Term Memory — all complex biological subsystems with similar trait profiles).

The supply chain exception is instructive: when a concept is genuinely domain-agnostic in its structural role, even embeddings detect cross-domain analogs. UHT’s advantage is that it finds these bridges for domain-specific concepts too — concepts that embeddings would never connect.

Action

HYP-046 is closed as confirmed (RES-CALIBRATIONRESULTS-047). The next productive direction is investigating whether hamming neighbors are semantically meaningful cross-domain analogs or merely hex-code noise. A follow-up hypothesis should test whether hamming neighbor pairs share identifiable structural roles (container, transformer, constraint) at rates above chance. The two active hypotheses remaining are HYP-036 (null-hex reclassification with enriched names) and HYP-042 (quadrant holism) — both may have closed duplicates and should be triaged before new hypothesis creation.