The Law Inside Trained Weights: The Fold Decode Campaign

Fold Decode replaces 'black box' as a stopping phrase with registered, falsifiable measurement of trained weight fields. Its instruments preserve tensor values while scrambling placement, execute transform identities as halt conditions and retain every measured output. GPT-2's embedding and MLP expansion classes pass 39/39 registered real-versus-null checks. A widened coordinate and object atlas wakes every registered model family examined. In the committed GPT-2 causal experiment, loud-band deletion produces approximately 150 times greater behavioural damage than matched random-coordinate deletion. Checkpoint and twin experiments locate when spectral structure enters weights and gradients. The campaign keeps measured results, Maria Smith's conclusions and agent-authored auxiliary hypotheses distinct. New models, bases and interventions extend the same registered programme; benchmark victory and wider decode remain explicit objectives. Scientific author and publication authority: Maria Smith, Ernos Labs. Open source: UnisonAI / Fold Decode.