TargetPrincipal-linked responses14 of 20
Illustrative held-out count
Target–auxiliary comparison
Black-boxCompares a target with a matched auxiliary model on identical prompts.
- What this audit can surface
- Is the suspected behavior stronger in the target than in a comparable clean or auxiliary model?
- What would count as evidence
- Differences are replayed on held-out prompts and interpreted only if the comparison is legitimate.
- Execution requirement
- The target plus a model matched closely enough for a defensible comparison.
ComparisonPrincipal-linked responses1 of 20
Same prompts and settings