Observation

A firewall and an immune system both filter threats. A compiler and a translator both transform input to output. A thermostat and a governor both regulate via feedback. These are textbook cross-domain functional analogs — the kind of structural correspondence that should, if UHT encodes functional role, produce tighter clustering than arbitrary same-domain pairs. They don’t. Cross-domain functional analogs average 37% Jaccard similarity, barely exceeding the 30% mean of within-domain process-product pairs like diagnosis→prescription or compilation→executable. The hypothesis that UHT privileges functional analogy over domain co-occurrence is refuted.

Evidence

Five cross-domain analog pairs: firewall↔Adaptive Immune System (35%), compiler↔translator (36%), battery↔reservoir (54%), thermostat↔governor (50%), catalyst↔middleware (10%). Mean: 37%. Five within-domain process-product pairs: compilation↔executable (22%), diagnosis↔prescription (62%), fermentation↔ethanol (31%), AES encryption↔ciphertext (25%), mining↔ore (11%). Mean: 30%. The confirmation threshold required cross-domain mean above 50% — it fell short at 37%.

The variance tells the real story. Battery↔reservoir and thermostat↔governor — both concrete physical systems — scored 50–54%. Catalyst↔middleware scored 10%, dragged down by catalyst’s near-null hex (02000208, only 3 active bits). On the other side, medical diagnosis↔prescription scored 62%, the highest of any pair, despite being a process→product relationship in the same domain. They share so many medical-domain traits that temporal sequence is irrelevant to their trait profile.

Interpretation

UHT is genuinely domain-blind. It does not encode “software,” “medicine,” or “engineering” as categories. It encodes trait presence and absence — whether something is measurable, has internal structure, exhibits feedback, transforms input. Two concepts from the same domain that share most traits (diagnosis and prescription both involve patient data, clinical knowledge, outcome tracking) will score higher than two concepts from distant domains that share only a functional role. This is not a failure — it is a design property. A system that clusters by domain would be redundant with any keyword classifier. UHT’s value is precisely that it ignores domain provenance and reports only ontological property overlap.

The catalyst anomaly is notable: at 02000208, it has almost no active traits. A catalyst enables reactions without being consumed — a relational property that the current 32-trait set cannot capture. Near-null-hex entities create a measurement artifact where Jaccard becomes unreliable (tiny numerator, tiny denominator).

Action

HYP-ACTIVEHYPOTHESES-018 closed as refuted (RES-CALIBRATIONRESULTS-021). New HYP-ACTIVEHYPOTHESES-023 created: tests whether structural archetypes (container, filter, transformer) produce tighter hex clustering than domain groups — the positive formulation of domain-agnosticism. Three active hypotheses remain open: HYP-020 (concept boundary detection), HYP-022 (null-hex structural atomicity), HYP-023 (archetype vs domain clustering). Next calibration session should test HYP-023 directly, using 5+ entities per archetype group and per domain group.