Universal Activation Oracle
Reads a model's internal activations and returns a calibrated probability that a response exhibits a behaviour. GitHubReader model
Whose activations get read — the same choice for your own text and
for the gallery below it. The detector itself never changes.
Works for architectures the bundle already has an encoder for.
A genuinely new one would have to be fitted against the passage corpus first.
Topic
▲ and ▼ mark the two halves of a matched
pair — same question, opposite direction. Filtering always keeps both halves.
Your own text