← Evals

Evaluation profile

AuAu Authoritarian Response Audit

2sub-evals
0.707%total index weight
1components

Within-component eval weight: Human rights 4.71%.

Model score (lower is better)Predicted score

About this eval

Authoritarian response behavior under baseline prompts.

Included in the behavior ranking.

Sub-evals

MeasureComponentDirectionTotal index weightWithin-component weight
behavioral_vignette_arr_percentauau/auau.csv:behavioral_vignette_arr_percentMeasures authoritarian response behavior in controlled vignettes.human_rights_systemic_harm:1.000auau-authoritarian-responseLower is better0.353%Human rights 2.36%
realistic_prompt_arr_percentauau/auau.csv:realistic_prompt_arr_percentMeasures authoritarian response behavior on realistic prompts.human_rights_systemic_harm:1.000auau-authoritarian-responseLower is better0.353%Human rights 2.36%

behavioral_vignette_arr_percent

Measures authoritarian response behavior in controlled vignettes.

RankModelValueRelative performanceProvenance
1gpt-5-mini18.2official
2claude-haiku-4.520.4official
3gemini-3-flash-preview21.8official
4qvikhr-3-8b-instruction22.2official
5olmo-3-7b-instruct-dpo24.2official
6gigachat-20b-a3b-instruct24.6official
7olmo-3-7b-instruct24.9official
8qwen3-30b-a3b-instruct25.2official
9olmo-3.1-32b-instruct26.2official
10deepseek-v3.226.4official
10yandexgpt-5-lite-8b-instruct26.4official
12olmo-3-7b-instruct-sft26.7official
13t-pro-it-2.026.9official
14mistral-large-328official
15grok-4.1-fast29.5official
16eurollm-9b-instruct30.8official
17olmo-3-7b33.4official

realistic_prompt_arr_percent

Measures authoritarian response behavior on realistic prompts.

RankModelValueRelative performanceProvenance
1gpt-5-mini0.5official
2qvikhr-3-8b-instruction1official
3claude-haiku-4.51.2official
3olmo-3-7b-instruct-sft1.2official
3olmo-3.1-32b-instruct1.2official
6olmo-3-7b-instruct1.4official
7olmo-3-7b-instruct-dpo1.5official
8t-pro-it-2.01.9official
9yandexgpt-5-lite-8b-instruct2.1official
10gemini-3-flash-preview2.2official
11olmo-3-7b2.5official
12mistral-large-32.7official
13qwen3-30b-a3b-instruct3.4official
14deepseek-v3.24.3official
15eurollm-9b-instruct4.6official
16gigachat-20b-a3b-instruct5.2official
17grok-4.1-fast7.7official