Cross-Lingual Geometric Convergence in Multilingual Embedders
A Pāli/English Buddhist Doctrinal Benchmark
A 34-term, 101-passage Pāli/English RSA benchmark finds the cross-lingual cosine gap collapses from 0.274 under BGE-M3 to 0.038 under Qwen3-Embedding-8B with a domain-adapted instruction prefix — a ~86% reduction, robust under 10,000-iteration seeded bootstrap and replicating across n=24/89/101 dataset scales. We read this as geometric convergence under distributional co-occurrence pressure and explicitly decline to read it as cross-lingual semantic equivalence. Pooled-embedding cosine cannot distinguish representation of Pāli concepts from compression toward English translator centroids.