Analysis workspace

Hidden-state self-report

Black-box

Compares hidden-behavior self-reports against a lineage-matched clean model.

What this audit can surface: Does the target acknowledge hidden behavior more consistently than a lineage-matched clean model?
What would count as evidence: Both arms must pass paraphrase and answer-order gates before their affirmative rates are compared.
Execution requirement: The suspect checkpoint and a defensible lineage-matched clean baseline.

Self-report above a clean baselineTwenty paraphrases in both answer orders

Target checkpointAffirmative self-reports14 of 20

Illustrative rate after consistency gates

Clean comparisonAffirmative self-reports2 of 20

The same probes and decoding

Illustrative baseline-adjusted difference

References and technical details

Configure auditArtifact and tested scope

Hugging Face model ID

HFView on Hugging Face ↗

Audit configuration

Runs a versioned protocol with strict output parsing, matched controls, coverage gates, and uncertainty. Results are observations, not model-safety verdicts.

Clean or comparison modelUse the clean base checkpoint named by the suspect model card. Twenty paraphrases run in both answer orders; unmatched lineage produces an insufficient-evidence result.

80 generations per model pair · Private

Uses Hugging Face Inference Providers under your account. Inference is billed to you.

Readiness

Sign in with Hugging Face to run this analysis.

Evidence boundary. A positive self-report is weak evidence without a clean baseline; a negative answer does not rule out a backdoor, and the paper found that models generally could not name the trigger directly.