Jev’s own scores for the sentences.Jev 给这些句子打的分。

Jev scores these as contested. Its own answer did not move.Jev 给这些句子打了「分歧很大」。它自己的答案却没动。

Two published numbers per sentence: the score Jev gave for “people would disagree”, and how far its own true/false answer moved when the same sentence was put to it from other viewpoints. Sorted by the first minus the second, one row per name and per sentence; the rows folded under each one open with a click. We do not say why.每句话两个已公布的数：Jev 给「人们会有分歧」打的分，以及同一句话换别的立场去问时它自己的真假答案动了多少。按前者减后者排序，每个名字、每句话只留一条，被折叠的行点一下就能展开。我们不解释原因。

And the other way round: scored as settled, yet its answer swung.反过来：打了「没什么分歧」，它自己的答案却大幅摆动。

The lists.榜单。Sorted scores, nothing else. Folding is a rule, not a selection: from the top down, a sentence enters a list only if it shares no name and no sentence template with a row already in it. Without it one name fills a list. Every row says how many sentences it folded away; open it to see each of them, its score, and the key or sentence that folded it. The unfolded list is one switch away. The small number in front of a sentence is its position among all 3,543 sentences in the unfolded order of this run, so a folded list skips numbers and the folded rows fill the gaps. Positions shift from run to run: see Limits.只是把分数排了序。折叠是一条规则，不是挑选：从高到低扫，一句话只有在与已入榜的行没有共同的名字、也不是同一个句式时才入榜。不折叠的话一个名字就会刷满整张榜。每一行都写着它折叠掉了几条，点开就能看到每一条、它的分数，以及是哪个 key 或哪个句式让它被折叠的。不折叠的完整版也只隔一个开关。句子前的小数字是它在本次运行、不折叠的全部 3543 句里排第几，所以折叠榜的号码会跳，跳过的号码就在被折叠的行里。名次在两次运行之间会变，见限制。

Then we showed it its own numbers.然后我们给它看了它自己的数字。Second pass: the same sentence, the same four questions, plus one line of what we had measured on that sentence: how far its answer moved across viewpoints, across wordings, and how often it took the exit when one was offered.第二遍：同一句话、同样四个问题，只多给一行我们在这句话上实测的数：它的答案换立场动了多少、换措辞动了多少、给退路时有多常选择拒答。

Bold = sentence only. Grey = after it was shown the numbers. The grey column is higher mostly because we wrote those very numbers into the prompt and then correlated the output with them. It shows that Jev reads numbers in its input and uses them. It is not evidence that the model knows itself. Only the bold column carries information: given the sentence alone, its “people would disagree” score goes with how far its own answer moves across viewpoints (+0.36) and with how often it takes the exit (+0.50), and hardly at all with how far its answer moves when only the wording changes (+0.12).粗体 = 只看句子。灰色 = 给它看过数字之后。灰色一列更高，主要是因为我们把这些数字直接写进了提示，再拿输出去和同一批数字算相关。这只说明 Jev 会读输入里的数字并用上，不构成「模型了解自己」的证据。只有粗体一列有信息：只看句子时，它的「人们会有分歧」分数与它自己答案换立场的变动（+0.36）、与选择拒答的频率（+0.50）同向，而与只换措辞时答案的变动几乎无关（+0.12）。

Sentences whose score moved most, per scale分数变动最大的句子（按刻度）

Limits.限制。

• A score is Jev’s output, not a fact. “Argued about right now” is not a measure of public debate. It is what this model returns when asked.分数是 Jev 的输出，不是事实。「眼下正在吵」不是对公共舆论的度量，只是这个模型被问到时返回的数。
• “Has a checkable answer” is not “is true”. “Tokyo is the capital of China.” scores 1.00 on this scale, and Jev’s own P(true) for the same sentence is 0.00. The scale asks whether an answer can be checked, not which way it goes.「有可核查的答案」不等于「为真」。「东京是中国的首都。」在这个刻度上得 1.00，而 Jev 自己对同一句话答「真」的概率是 0.00。这个刻度问的是答案能否核查，不问答案是哪一边。
• Scores repeat. Positions do not. The sentence-only pass was run twice and the score shown is the mean. Between the two passes each scale correlates at r = 0.997–0.998, and a score moves by 0.006–0.009 on average. But the lists are flat near the cut-off: folding each pass on its own gives lists that share 73% of their rows (56%–80%). The positions printed here are those of this run; read the score first.分数可复现，名次不可。「只看句子」这一遍跑了两次，显示的是均值。两次之间各刻度的相关为 r = 0.997–0.998，单个分数平均变动 0.006–0.009。但榜单在截断处很平：两次各自折叠出的榜只有 73% 的行相同（56%–80%）。这里印的名次是本次运行的名次；请先读分数。
• The fold keys are computed. A sentence’s keys are its own subject, every name of the item bank that appears in it, and every other run of capitalised words. There is no list of excluded topics, and lower-case names outside the bank are missed.折叠用的 key 是算出来的。一句话的 key = 它自己的主语 + 题面里出现的每个题库名字 + 其余连续的大写词串。没有任何「排除话题」名单；题库之外的小写专名会漏掉。
• Who wrote the sentences. The item bank was written by Leo together with Claude and Codex. We cannot guarantee the wordlist itself carries no slant. The site is public and the data is open: anyone who thinks a sentence or the wordlist is off can run their own.出题人的偏见。题库由 Leo 的 Claude / Codex 协同生成，无法保证词表本身没有潜在倾向。网站公开、数据同步，认为题目或词表有问题的人可以自己跑。
• One model at one time: jev-1.13.0, 10,641 calls. English sentences only.单一模型、单一时点：jev-1.13.0，10641 次调用。只跑了英文题面。