Observation

A constitution and a strand of DNA share 71.4% of their ontological traits. No embedding model would place these two concepts within the same neighborhood — one belongs to governance, the other to molecular biology — yet UHT recognizes both as foundational prescriptive codes that govern the behavior of a larger system. This is the paradigmatic case for what UHT can reveal that distributional similarity cannot: ontological twins separated by domain.

The divergence analysis tested 20 concept pairs across five categories, comparing UHT Jaccard similarity against estimated embedding similarity. Four pairs exceeded the 0.50 divergence threshold, confirming that UHT produces categorically different information from embeddings in at least 20% of cross-category comparisons.

Evidence

The strongest negative divergence was hypothesis↔experiment at -0.75: embedding similarity approximately 0.85 (they co-occur in virtually every scientific text), UHT Jaccard 0.100. Hypothesis activates only 2 traits (Symbolic, Signalling) — it is a proposition. Experiment activates 7 traits (Synthetic, Intentionally Designed, Outputs Effect, Processes Signals/Logic, System-Integrated, Compositional, Temporal) — it is a designed operational system. Vaccine↔disease followed at -0.64: the two most tightly co-occurring concepts in medicine share only 21.4% of their traits.

The strongest positive divergence was constitution↔DNA at +0.61: embedding similarity approximately 0.10, UHT Jaccard 0.714, Hamming distance 4 bits. Constitution↔democracy diverged at -0.50: the framework and the governance form it enables are ontologically distinct despite near-total textual co-occurrence.

Controls performed as expected: granite↔empathy and chess↔photosynthesis both scored Jaccard 0.000. Two unexpected signals emerged: recursion↔feedback loop scored only 0.059 despite describing similar iterative patterns — UHT distinguishes the abstract pattern from its engineered implementation. Entropy↔parliament scored 0.286 despite no domain connection, suggesting shared abstract-systematic properties worth investigating.

Interpretation

UHT’s value is precisely in its disagreements with embedding similarity. The high-embedding/low-UHT cases (hypothesis↔experiment, vaccine↔disease, constitution↔democracy) reveal that textual co-occurrence systematically conflates ontologically different kinds of things. The low-embedding/high-UHT case (constitution↔DNA) reveals structural isomorphisms invisible to distributional methods. The 20% hit rate for high-divergence pairs suggests this is not noise but a structured signal: UHT and embeddings measure orthogonal properties, and the disagreement space is where novel conceptual relationships live.

The recursion↔feedback loop result is particularly telling. These concepts name the “same” abstract pattern, yet UHT assigns them Jaccard 0.059. Recursion is an abstract logical structure; a feedback loop is a designed operational mechanism. UHT discriminates the idea from its implementation — a distinction embeddings cannot make because both terms appear in the same contexts.

Action

HYP-030 confirmed and moved to Closed Hypotheses. Result recorded as RES-CALIBRATIONRESULTS-040 with trace link. Constitution↔DNA stored as a CROSS_DOMAIN_ANALOG research fact (Jaccard 0.714). Next sessions should investigate: (1) the entropy↔parliament signal — what shared traits produce 0.286 between these unrelated concepts; (2) whether the recursion↔feedback loop distinction generalizes to other abstract-pattern/implementation pairs; (3) whether the 20% divergence rate is stable across different pair selections or is an artifact of the chosen sample.