Templatic Morphology as Decodable Geometry, and Abjad Letter-Values as an Attention Probe, in Classical Arabic Transformer Models

A pre-registered cross-linguistic study of templatic Semitic morphology and letter-value structure in classical-Arabic and classical-Hebrew transformer encoders (CAMeLBERT-ca, AlephBERT). We report that the Arabic morphological pattern (wazn) is encoded as geometrically separable, linearly decodable structure, and that cross-root word pairs with proximate Abjad values receive higher mutual attention than Abjad-distal pairs (partial ρ = −0.0558, n = 2,322,977; 1,000-permutation p = 0.026). A pre-registered Hebrew control satisfies the morphologically-general outcome, and a frequency-matched null locates the Mashriqi-specific component at the level of frequency-tier alignment. Code and locked result artifacts are released at the linked repository.