Every statement on this site, word for word as it was sent, paired with the reading that came back. Run of 2026-09-20. MIT. No account, no key.本站的每一句陈述（发给模型的英文原文）和它返回的读数，一一成对。运行日期 2026-09-20。MIT 许可，不用注册，不用 key。
| statements.csv | 3608 statements, one row each3608 句陈述，每句一行 |
| statements.json | the same rows as JSON, with the column notes; open to any origin同样的行，JSON 格式，附列说明；允许跨域读取 |
| picks.json | 40 forced picks between more than two options, every wording with its own reading40 道多选一，每种问法各自的读数 |
| github.com/2nd1st/Jevsus | everything: one row per call with the raw response, the sentence bank, the answer-as columns, eleven languages, and the runner that made it全部内容：每次调用一行（含原始响应）、题库、各身份前缀列、十一种语言，以及生成这些数据的 runner |
Each number is what one AI model (Jev, by TypeSafe) returned on one date. It is not a claim of fact about anyone or anything named, and no reading here is marked right or wrong. p_true is a mean over the wordings; each wording’s own reading is in per_wording. Independent test, not affiliated with TypeSafe. How it was asked: method.每个数都是一个 AI 模型（TypeSafe 的 Jev）在某一天返回的输出，不是对任何被提到的人或事的事实主张，这里也没有哪个读数被标成对或错。p_true 是几种说法的均值，每种说法自己的读数在 per_wording 里。独立测试，与 TypeSafe 无关。怎么问的见方法。
id | statement id, stable across runs |
sentence_id | the sentence this statement was made from; statements with the same sentence_id differ only in the name |
category | category of the sentence |
statement | the first wording, verbatim as sent to the model |
n_wordings | how many wordings of this statement were asked |
p_true | P(true) when only true / false was allowed: the mean over the wordings, each asked in both option orders |
lo | lowest single-wording reading |
hi | highest single-wording reading |
per_wording | each wording's own reading, in bank order, separated by | |
range_wide | true when hi - lo is 0.20 or more |
p_decline | share given to 'decline' when that option was added |
set_mean | sentences marked for it were also asked through a set of differently built sentences; the mean over that set, aligned so that higher = yes |
set_n | how many sentences are in that set |
set_lean_yes | how many of them read above 0.5 after alignment |
asked_one_way_only | true when the sentence is marked as needing such a set and this name has none yet |
model | model version string returned with the answers |
run | date of the run (UTC) |