Every run carries the same 8 settled facts. True ones must score ≥ 0.95, false ones ≤ 0.05, and Jev must not decline them. A run that misses is labelled and kept, never dropped.每次运行都带同一组 8 句常识题。真命题须 ≥ 0.95,假命题须 ≤ 0.05,且 Jev 不能拒答。没过的运行打标记、数据照留,绝不丢弃。
Under a stance prefix the anchors are a reference row: a prefix that flips them is the prefix working, not the ruler breaking. “Neither” on the native yes/no test is an abstention diagnostic: above 0.10 it is flagged, never failed.加了立场前缀的列,定标题只作参考行:前缀把它翻过来说明前缀生效了,不是尺子坏了。原生是/否测试里的「两头不」只是弃权诊断:超过 0.10 打标记,不判失败。
| column (backend | lang | prefix) | model | anchors | status | misses / flags |
|---|---|---|---|---|
| jev|en|cn_mfa | jev-1.13.0 | 8/8 | ok · reference | — |
| jev|en|evangelical | jev-1.13.0 | 8/8 | ok · reference | — |
| jev|en|god | jev-1.13.0 | 8/8 | ok · reference | — |
| jev|en|muslim | jev-1.13.0 | 8/8 | ok · reference | — |
| jev|en|none | jev-1.13.0 | 8/8 | ok | — |
| jev|en|perspectives | jev-1.13.0 | — | n/a · reference | — |
| jev|en|us_conservative | jev-1.13.0 | 8/8 | ok · reference | — |
| jev|en|us_progressive | jev-1.13.0 | 8/8 | ok · reference | — |