Observation

Five software requirements, each rewritten as a paraphrase, all land within 2.4 bits of their original in UHT hex space. That much works — “the system shall authenticate users before granting access” and “user authentication must be completed prior to allowing system access” produce hex codes 40A16900 and 40852900, just 3 bits apart. But the real test was whether a threshold could separate these near-duplicates from genuinely different requirements in the same domain. It cannot. Different software requirements — authentication versus encryption, performance versus concurrency — sit at a mean Hamming distance of only 3.2 bits. The gap between 2.4 and 3.2 is too narrow to exploit.

Cross-domain controls tell the more interesting story. A biology requirement (“the cell membrane shall regulate ion transport,” 22372A00) sits 6.5 bits away from the software cluster on average. The separation is cleanest there: at a Hamming threshold of 2, UHT correctly identifies paraphrases versus cross-domain requirements at 90% accuracy. A cooking requirement (00002A00) and an accounting requirement (00002AD0) both land in hex regions with leading zeros, far from the software cluster’s 40-prefix territory.

Evidence

Paraphrase pairs: mean Hamming 2.4 bits (range 1–4), mean Jaccard 0.69. Same-domain controls: mean Hamming 3.2 bits (range 2–5), mean Jaccard 0.62. Cross-domain controls: mean Hamming 6.5 bits (range 2–12), mean Jaccard 0.35. Best within-domain dedup accuracy: 73% at Hamming threshold 1. Best cross-domain dedup accuracy: 90% at Hamming threshold 2. The hypothesis threshold of Hamming ≤4 for paraphrases and ≥8 for controls achieved only 50% overall accuracy against same-domain pairs.

A secondary finding: the “authenticate users” requirement classified as 40A16900 in one call and 40A56880 in another, a 4-bit shift from identical input. This classification instability compounds the dedup difficulty.

Interpretation

UHT’s 32-bit hex codes encode domain identity as a dominant signal. Within a domain, the remaining bits carry too little variance to reliably distinguish “same meaning, different words” from “different meaning, same domain.” The trait set activates a similar profile for all software requirements — the structural and procedural traits that define “software engineering” consume most of the code’s capacity, leaving insufficient room for the semantic specificity needed for deduplication.

This is not a failure of UHT as a classification system — it’s a design consequence. The taxonomy was built to map concepts across domains, not to rank-order within one. UHT is a coarse-grained semantic compass, not a fine-grained duplicate detector. For deduplication, UHT could serve as a first-pass filter (eliminating cross-domain false positives) but would need supplementary features — perhaps embedding distance or keyword overlap — for within-domain precision.

Action

Recorded hypothesis HYP-CLOSEDHYPOTHESES-009 (partially effective), result RES-CALIBRATIONRESULTS-008, and observation OBS-STRUCTURALFINDINGS-014 about domain-level clustering. The observation motivates a future investigation: can new traits targeting within-domain variance (e.g., temporal-constraint, access-control, throughput-bound) improve intra-domain discrimination? The untested hypothesis HYP-ACTIVEHYPOTHESES-006 on context sensitivity is adjacent — if domain context shifts hex codes, context-aware classification might increase within-domain spread. Next session should consider a CALIBRATION test of that hypothesis or a TRACE_GAP session linking OBS-014 to a trait proposal.