Stability-Aware Auditing of Latent Regions in Sentence Embedding Spaces: Guarding Against Granularity Collapse in Unsupervised Text Analytics

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Sentence-embedding clustering is often interpreted from a single partition, although a reproducible solution can still be excessively coarse, cardinality-variable, or representation-dependent. We present a stability-aware audit framework that separates label-free operating-point selection from post hoc taxonomy evaluation and explicitly distinguishes the original frozen confirmatory execution from separately frozen additional robustness analyses. On a pinned Banking77 test revision (N=3076) and CLINC150+OOS (N=5500), the original anti-collapse rule selected seven-region MeanShift operating points with disjoint-random-stream confirmation ARIs of 0.9565±0.0510 and 0.9347±0.0327, respectively. Additional paired refits showed that rejected coarse candidates can remain more stable than the selected solutions, and this stability–granularity inversion persisted under ARI, AMI, NMI, V-measure, and Fowlkes–Mallows similarity. The additional audits also showed that high membership similarity can coexist with refit-cardinality variation, that representation and broader pipeline perturbations materially affect the discovered structure, and that density-based stability must be interpreted jointly with coverage and cardinality. These results support a conservative conclusion: perturbation stability is evidence of reproducibility under a stated protocol, not proof of an appropriate resolution, semantic validity, or taxonomy recovery.