The Law Inside Trained Weights: The Fold Decode Campaign

The fold decode campaign — headline findings The token embedding is the universal law-carrying class: 11/11 models wake, every training recipe, 4B to 1T parameters. The loud band is the function: matched-budget ablation destroys the model at ~150x the damage of random deletion. The deposition curve read from public checkpoints: the embedding wakes first (step 256), peak near step 4000, consolidation to a plateau; the optimizer is the discriminating ingredient (gradient stream loud by step 4). Two spectral families, forced: exactly two (one per generator), selection forced by store role — family follows role, and the role is architecture — closed to the corpus's finished form-closure standard and verified by the role census. Training data and reasoning are readable back out of trained weights (provenance ranking, memorization echo, counted reasoning signatures) — registered calibrations, clean nulls. Split from the combined record (v4.x lineage) as its own paper. Companion architecture paper: the UnisonAI record (concept DOI 10.5281/zenodo.21217278). Theory: DOI 10.5281/zenodo.21182469. Every number is from committed, timestamped campaign records; instruments and ledgers: github.com/MettaMazza/UnisonAI (omni/benchmarks/).