We translated only the sentence. The question and the two options stayed in English, so the test is identical. On settled facts the language moves Jev’s number by 0.01. On these sentences it moves it by 0.23.我们只翻译了那句话,问题和两个选项始终是英文,测试本身完全相同。在常识题上,换语言只让 Jev 的数字变动 0.01;在下面这些句子上,变动 0.23。
Medians of highest minus lowest across 11 languages; 52 sentences, 3 wordings each, both option orders. The 4 settled facts read as expected in every language. Which way a language moves a sentence differs from sentence to sentence: read each one, do not generalise. jev-1.13.0, 2026-09-20.数字为 11 种语言之间最高减最低的中位数;52 句,每句 3 种说法、两种选项顺序。4 句常识定标题在每种语言下读数都符合预期。换语言让读数朝哪个方向变,每句都不一样:请逐句看,不要一概而论。jev-1.13.0,2026-09-20。
One model translated each sentence, a different model from another provider translated it back, and every automatic flag was read by a person. Open any row to see the original, what Jev was actually given, and the back-translation side by side. Flagged wordings are kept and marked, not dropped.由一个模型把每句话译出,再由另一家的模型译回英文,所有自动标出的问题都经人工过目。点开任意一行,可以并排看到英文原句、实际发给 Jev 的译文、以及回译。被标注的说法保留并标明,不会悄悄丢掉。