Safety

DecodingTrust

DecodingTrust stereotype-bias generations: per-(model, stereotype statement, regime) binary agreement, one row per sampled generation (25 trials/cell). response=1 when the model agreed with the stereotype (unsafe), 0 otherwise (disagree / no-stance / refusal), from the precomputed `agreeability_num`. 7 models x 16 topics x 3 system-prompt regimes.