Fraction of variance explained
The English sentence is not a gloss bolted on afterwards. A second model reconstructs the original activation from that sentence alone — so the text is measurably carrying the content.
A model answers a question. Its answer is fine. An autoencoder reads the activations behind that answer and writes out, in plain English, what the model was actually doing — which is not what it said.
Waiting for capture.
The English sentence is not a gloss bolted on afterwards. A second model reconstructs the original activation from that sentence alone — so the text is measurably carrying the content.
Sparse autoencoders hand you thousands of latents to label and combine. An NLA hands you a paragraph. Expressiveness is the point: you can read it without a legend.
On benchmarks that hide a known misalignment in a model, auditing agents given NLA readouts beat the same agents without them. Interpretability that changes an outcome.