Skip to content

External activation-probe classifier contract

ActivationProbeClassifier is a mean-pooled logistic probe. It accepts either [seq_len, hidden_dim] activations or one [hidden_dim] summary vector and returns score, label, and confidence, all independent of the MAC hub.

Checkpoints are JSON with weights, bias, threshold, and optional positive and negative labels. A missing checkpoint constructs a disabled neutral probe (score=0.5, label=disabled, confidence=0). Dimension and finite-value checks fail closed inside the classifier, while the worker integration catches those failures because the overall audit is advisory.

This probe classifies activations supplied by its caller. It does not compute a Jacobian lens, acquire activations from hosted models, infer hidden states from proxy traffic, or expose the internal reasoning of a different model. It is not an intent oracle. Its checkpoint must be trained separately from its calibration data, versioned with the instrumented model and layer identity, and treated as a diagnostic signal only.