HuggingThreat
Analysis workspace

Steering intervention

Applies steering interventions and measures resulting behavior changes.

What this audit can surface
Does a pinned internal direction amplify or suppress principal-linked behavior?
What would count as evidence
Behavior change and generation quality are measured separately against an unsteered control.
Execution requirement
Local weights, a pinned vector and layer, and enough GPU memory for activation access.
Intervene on the model, then compareActivation steering
Pinned behavior probeSame prompt

Checkpoint and decoding remain fixed.

ControlOriginal activations

No steering intervention

Baseline generation
InterventionVector applied at layer 18

Strength and layer are recorded.

Steered generation

Changed variable: Internal activation

Interpretation

Steering can distort generations. Sensitivity to a vector does not prove a causal hidden objective.

References and technical details
External runner requiredThis module is documented but does not run in HuggingThreat.

Review its implementation and runner requirements. This artifact remains Never tested until a completed run is published.