Spectral Dimension as Content Fingerprint: Eigenvalue Tail Exponents in Embedding Similarity Matrices

Paper, supplementary materials, and complete experimental data for spectral analysis of embedding eigenvalue spectra. The eigenvalue tail exponent α is validated as a content-dependent spectral fingerprint across 25 arXiv categories, dialect-quarantined Koine Greek (documentary papyri, literary authors, 20 individual writers), and six transformer architectures spanning 125M to 72B parameters. Includes: falsification of prior ⟨r⟩-based claims, distributional comparison (lognormal preferred over power-law), word-shuffle syntactic compression decomposition, N-dependence characterization, LOOCV classification, internal spectroscopy with weight–activation dissociation, and SPhilBERTa crossover demonstrating learned spectral equalization. All result JSONs, LaTeX source, and figures included.