NLAs convert model activations into readable text by training an explainer and checking that a second model can rebuild the activation from it. Released code and checkpoints let you inspect open models and audit hidden behavior such as test awareness.
