The 99-Names Coordinate System: A Fitted Relational Basis for Classical Arabic in CAMeLBERT-ca

A methodology and resource for incorporating a documented scholarly relational structure as a fitted coordinate basis on the contextual embedding space of CAMeLBERT-ca. Distributional similarity tools (cosine, k-NN, contextual embedding distance) cannot reliably resolve fine-grained relational distinctions in classical Arabic vocabulary: near-synonyms, speech-act variants, and categorically distinct emotions collapse into a narrow cosine band (0.865–0.937) that carries no information about the kind of relationship each pair has. We use the 99 Names of God — a closed, well-attested set with hierarchical, root-family, morphological, and documented opposition relations as a coordinate-defining basis, treating it as a methodological instrument rather than a theological commitment. Arabic terms are projected into the basis via CAMeLBERT-ca embeddings followed by Poincaré-disk reduction, yielding a position, hierarchical tier, attractor profile, and repellor profile for each term. Six empirical case studies demonstrate distinctions the cosine baseline cannot make. The release includes the exact 99-entry, 30-field basis source used by the deployed engine (with field-level evidential-status labels), the fitted Poincaré coordinates, machine-readable metadata, a schema validator, the engine code, and a live MCP server exposing five operations. The methodology generalises to any documented relational structure in any language given an appropriate contextual embedding model.