Two traces in. One audit out.

Pythia compares two traces. The difference is the audit.

We run the base model and the fine-tune on the same prompts, decompose their internal activations into sparse features, and surface every feature whose firing changed. Concept localization at the layer × feature level.