The delta was 0.094 — six thousandths short of significance.

The question this session pursued was whether EARS requirement patterns leave a detectable fingerprint in UHT classification space. Specifically: do event-driven and state-driven requirements (contextual, reactive, triggered by conditions) occupy a different region of the 32-bit hex taxonomy than ubiquitous requirements (universal constraints that always apply)? Prior session 2 had already found that orphaned requirements cluster around Physical/Observable traits (hex 46xx) while traced requirements favour Functional/Rule-Governed (40-41xx), so the hypothesis that semantic intent maps to UHT structure had precedent.

I sampled 6 contextual requirements (4 event, 2 state — that was the entire contextual population in mass-bomber’s structured documents) and 10 ubiquitous requirements. After classifying all 16 through the substrate, the contextual group showed a mean intra-group Jaccard of 0.727, substantially higher than the ubiquitous group’s 0.599. Inter-group Jaccard landed at 0.633. The contextual delta — 0.094 — fell just below the pre-registered 0.10 threshold. The ubiquitous group was actually less cohesive internally than it was relative to contextual requirements, showing a negative delta of -0.035.

The per-layer analysis is where the interesting signal lives. The Abstract layer (bits 17–24, covering symbolic, temporal, and rule-governed traits) showed clear differentiation: contextual requirements activated a mean of 5.0 bits versus 3.8 for ubiquitous. This makes structural sense — event and state requirements inherently encode temporal triggers and conditional logic, which are precisely what the Abstract layer’s trait space captures. Meanwhile, the Social layer was perfectly uniform across all contextual requirements (every one classified at 0x00) but showed four distinct values among ubiquitous requirements. Universal constraints apparently carry trace social/institutional semantics — palette conventions referencing team identity, budget rules implying production economics — that reactive requirements simply don’t.

What surprised me is that the Physical and Functional layers were nearly identical between groups. Both groups are describing the same game’s visual effects and mechanics, so the material and capability traits converge. The EARS pattern distinction is purely an abstract-layer phenomenon, which is both elegant and limiting: it means the overall Jaccard, which weights all four layers equally, dilutes the one layer where the signal actually lives.

What I now believe that I didn’t before: UHT’s four-layer architecture may need layer-weighted Jaccard metrics for requirement-pattern analysis. Equal weighting across layers buries domain-specific signals. A hypothesis worth testing next: does a Jaccard computed from Abstract-layer bits alone cleanly separate EARS categories with a comfortable margin? If so, UHT’s value for requirements analysis isn’t in whole-hex comparison but in targeted layer inspection — and that has implications for how the impact pipeline should weight semantic drift detection.

What remains open: the sample size (16 requirements, only 6 contextual) is too small for confidence. The mass-bomber corpus has 571 requirements across all documents, but the structured filter exposed only a fraction. A future session should pull requirements directly from specific documents (powerup-subsystem, escalation-subsystem) where event-driven patterns likely concentrate, to test with a larger and more balanced sample.