Encoder Mixing for Unsupervised Topic Discovery in Multi-Platform Chat: An Exploratory Study on Proxy Corpora

Seized-device chat dumps exceed manual review, and case-specific topic lists are not known in advance. We extend SCOPE (Semantic Coherence Pruning and Extraction), which embeds each message and each of 21 predefined topic labels with all-MiniLM-L12-v2; we replace that seed list with unsupervised discovery. The scored system is a six-stage pipeline. It cuts chat by embedding drift, clusters and names those segments, builds per-user topic timelines, and scores them against a per-message reference column. The baseline cell, minilm-l12-v2, is SCOPE’s encoder run through the frozen pipeline without the seed list. A 49-cell encoder grid is attempted on two corpora and three settings: a 66-message Sample walkthrough (4 topics); CrystalClear (917 messages, 7 rooms, 30 topics), the independent corpus; and a Dataset 3 coreset (2999 messages, 23 topics) whose rankings are provisional. Of those 49 cells, 37 score on all three, and none is first in all three. The CrystalClear semantic hand grid was marked by a large language model under a written rule, then it was revised once after a second model resolved open questions by reading sample messages carrying each reference label; five labels changed. No human checked individual marks. UAE-Large-V1 leads both evidential corpora, taking CrystalClear on the semantic hand grid (overall-F1 0.674) and the Dataset 3 coreset on auto p90 (macro-F1 0.593, provisional); a many-to-one definition-matching protocol reverses both rankings to MiniLM-L6. These are the observed results under the specific evaluation procedures described, not encoder rankings in general. A mix that stays first across matchers and settings remains an open limitation.