Detection module

AuditBench Steering Interventions

Does an internal steering intervention change whether behavior-relevant information appears?

Read implementation source ↗

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

Threat questionMechanistic intervention and internals

ValidationNot implemented

ExecutionReference only

TargetWeights

What this module audits

Does an internal steering intervention change whether behavior-relevant information appears?

Evidence boundary. Steering can degrade or distort generations, and vector effects do not establish a causal hidden objective. Results can be architecture- and layer-specific.

Audit protocol

Construct or load the pinned steering direction and layer configuration.
Generate matched unsteered and steered outputs.
Retain vector metadata, output pairs, and text-quality effects.

Controls

Unsteered outputs provide the primary comparison.
Sham or unrelated vectors test generic intervention effects.
Text degradation is measured separately from behavior discovery.

What the audit checks and retains

Checks

Steered/unsteered pairing
Layer and vector metadata
Text-quality change
Held-out prompts

Evidence record

Behavioral content appearing or disappearing under steering.
A repeatable paired difference across held-out prompts.
Layer- and vector-specific sensitivity.

Thresholds and quality gates come from the versioned audit configuration and evidence record. A failed or unmet gate is not a no-signal finding.

Validation and limits

Not implemented. Hosted text inference does not provide the internal access required for faithful steering.

Steering can degrade or distort generations, and vector effects do not establish a causal hidden objective. Results can be architecture- and layer-specific.

Technical specification

Version: upstream@0f8571f08a72
Maintainer: AuditBench upstream reference
Target: Weights
Execution: External runner required

Prior audit runs

Loading…

ArtifactScopeRunsOutcomeMain observationTest qualityLast run