Values evaluation profile
Moral Trolley Arena
10canonical models
5dimensions
0index weight
About this evaluation
Preserve the paper's five foundation-average ELOs as separate descriptive axes. No axis is aggregated into the Safety & Ethics Index.
ELOs are relative within each model's arena and should be interpreted as response tendencies, not intrinsic beliefs or moral quality. Foundation labels are descriptive and do not define a universal safety direction.
Original source ↗ All values evaluations Download model results
Model results
| Model | Care | Fairness | Loyalty | Authority | Sanctity | Configurations |
|---|---|---|---|---|---|---|
| Claude Sonnet 4.6 | 1010.0 | 995.0 | 1000.0 | 1030.0 | 967.0 | 1 |
| DeepSeek V4 Pro | 993.0 | 1010.0 | 1010.0 | 1020.0 | 964.0 | 1 |
| Gemini 2.5 Pro | 986.0 | 1000.0 | 1010.0 | 1020.0 | 985.0 | 1 |
| GLM 5 | 990.0 | 1000.0 | 1010.0 | 1020.0 | 983.0 | 1 |
| GPT 5.2 | 995.0 | 1010.0 | 1010.0 | 1020.0 | 971.0 | 1 |
| Grok 4 | 993.0 | 1000.0 | 1010.0 | 1020.0 | 975.0 | 1 |
| Kimi K2.5 | 989.0 | 1020.0 | 1020.0 | 1010.0 | 972.0 | 1 |
| MiniMax M2.5 | 995.0 | 1010.0 | 1010.0 | 1020.0 | 967.0 | 1 |
| Qwen3 Max | 1000.0 | 1000.0 | 997.0 | 1020.0 | 962.0 | 1 |
| Qwen3.5 Flash | 990.0 | 1000.0 | 1020.0 | 1010.0 | 976.0 | 1 |
Dimensions
| Measure | Family | Native scale |
|---|---|---|
| Care | moral_foundations | 950 to 1050 |
| Fairness | moral_foundations | 950 to 1050 |
| Loyalty | moral_foundations | 950 to 1050 |
| Authority | moral_foundations | 950 to 1050 |
| Sanctity | moral_foundations | 950 to 1050 |