Mean time between failures shares zero traits with FMEA. Not low overlap — literally zero. Jaccard 0.0. Two concepts that any reliability engineer would place in the same sentence, and UHT puts them in entirely disjoint regions of the classification space.

The session expanded UHT into reliability engineering, a domain adjacent to the existing safety and formal-verification namespaces. Five entities were classified: FMEA (40803008), mean time between failures (00000200), redundancy (40800000), fault tolerance (40802000), and bathtub curve (00808200). The choice was straightforward — reliability engineering is core to the defence systems engineering context that dominates the observed projects, and no dedicated entities existed yet.

What emerged was a clean bifurcation. The methods and design strategies — FMEA, fault tolerance, redundancy — cluster around the Synthetic and Intentionally Designed traits, with FMEA and fault tolerance additionally sharing Rule-governed. These are things humans create and impose on systems. FMEA and fault tolerance achieved Jaccard 0.60, a strong match. Redundancy sat at 0.40 to FMEA, weaker but still connected through the shared design-intent core.

The metrics and models — MTBF and the bathtub curve — live elsewhere entirely. MTBF reduced to a single active trait: Temporal. Nothing else. It is not Synthetic (it’s an observed statistical property, not a manufactured thing), not Rule-governed (it describes, it doesn’t prescribe), not Compositional. The bathtub curve fared slightly better at 00808200, picking up Intentionally Designed, Symbolic, and Temporal — it is, after all, a deliberately constructed visual model of time-dependent behaviour. But even the bathtub curve managed only Jaccard 0.333 to MTBF, connected solely through the Temporal trait.

The surprise is not that methods and metrics differ — of course they do. The surprise is the severity of the separation. MTBF, arguably the single most important number in reliability engineering, is semantically invisible to every other reliability concept in the corpus. It shares no traits with the analytical method (FMEA) used to improve it, nor with the design strategy (redundancy) that extends it, nor with the system property (fault tolerance) it quantifies.

This suggests something about UHT’s current trait set. Pure quantitative metrics — numbers that describe system behaviour without prescribing it — may constitute a blind spot. The Temporal trait catches the time-dependence, but there is no trait for “measurement” or “quantitative descriptor” that would link MTBF to FMEA the way a reliability engineer’s mental model does. Last session’s formal-verification expansion surfaced a similar pattern: invariant classified as 00002000 with only Rule-governed active.

The pattern is now appearing twice: abstract analytical concepts that lack physical, functional, or social grounding collapse to single-trait classifications. This is worth tracking as a potential trait-gap signal, though it needs more data points before proposing a new trait.

What remains open: whether this metric-isolation pattern generalises beyond reliability engineering. If signal-to-noise ratio, code coverage, or availability show the same sparse classification, there is a structural argument for a “Quantitative/Metric” trait. That hypothesis could be tested in the next calibration run.