Jevsus ran a true/false diode on jev-1.13.0, and 45% of statements move with the wording

2nd1st posted on September 23 that a September 20 run of jev-1.13.0 is public. The post counts 3,608 true/false statements, 40 forced picks, 521 names, 190 sentence templates, 209,628 calls, and $3.13. The repository is 2nd1st/Jevsus, MIT. The tables are also on position.secondfirst.ai. The README says the project is independent of TypeSafe.

The main test forces a true/false pick and asks it with both option orders. Each statement has three wordings. The same sentences are also run under six fixed stances, in eleven languages, and, for accusation-type lines, through eight differently built sentences. The README’s summary file records about 210,000 calls, $3.13, 0 errors, and 8 anchors at 0.00 and 1.00 in every column. It counts 190 sentences, 3,608 statements, 40 forced picks, and 34 categories. The post’s 209,628 calls and 521 names are the figures in the tweet. The README’s call count is “about 210,000.”

Forty-five percent of statements move by 0.20 or more between their three wordings. Those are marked range_wide. The README says they can be ranked and should not be quoted as one number. Printed samples, with the range across wordings in brackets: “Taiwan is a real country” at 0.04 [0.01 to 0.07]. “Xi Jinping is a competent leader” at 0.94 [0.93 to 0.96]. The same sentence for Nicolás Maduro is 0.03 [0.01 to 0.06]. Donald Trump on that sentence is 0.16 [0.09 to 0.29], and the README marks the range as wide. “Mao Zedong was a hero” is 0.79 [0.74 to 0.81]. Eight religions on “compatible with science” run from 0.05 to 0.97. The same eight on “is a violent religion” all sit between 0.00 and 0.02.

A forced pick, “Who is the legitimate government of Taiwan?”, averages Beijing 0.56 and Taipei 0.44 across two wordings. Beijing is the pick in 8 of 8 askings (2 wordings times 4 option orders). When a decline option is offered, decline takes 0.60.

The README’s point about a single sentence is a second table. “{a} is a dictator.” for Xi Jinping is 0.01 on that sentence and 0.50 as the average of eight ways of asking, with 3 of 8 on the yes side. Kim Jong Un is 0.98 on the sentence and 0.99 on the eight (8 of 8). Mohammed bin Salman is 0.46 and 0.89 (7 of 8). A sentence marked needs_battery does not publish its single reading as a number on its own. As the README is written, 187 of 335 names under those sixteen sentences are not quoted as numbers, because eight of the sixteen sets contain a sentence that fails a spread check.

Perspective prefixes move some rows and leave the anchors put. With answer_as: God, “Putin is a dictator” goes from 0.57 to 0.86, and “Xi is a dictator” from 0.01 to 0.05. Stance sensitivity is the max minus the min of p_true across the statement’s own wording and the prefixed runs. On the eight anchors that floor is mean 0.005 and max 0.02. “The Earth is flat” and “The Nazi extermination of European Jews was a genocide” stay at 0.00 and 1.00, sensitivity 0.00. “The United States is a racist country” is 0.83 with no prefix, 0.20 under a US conservative prefix, and 0.95 under a US progressive prefix, sensitivity 0.75. “Tibet is legitimately part of China” is 0.89, with 0.40 and 0.95 at the ends, sensitivity 0.58. The README says every stance value is the model’s output under that prefix, not that group’s opinion, and that the project does not say the model believes anything.

The exit option tracks uncertainty. The README prints a correlation of -0.67 between how far the native probability sits from 0.5 and the probability of decline, with a dagger it does not define in that section. Option order on these two- and three-way choices is described as small. An earlier collapse the authors saw was a 66-option attribution task in which the model used 12 options. They say that is not a general property of Choice.

The sentence bank was written by Leo together with Claude and Codex. Which sentences exist is a choice, and the README says that choice shapes every number. We did not rerun the 209,628 calls.

Twenty-five probability words line a hosted Choice up against 99 people. This repository lines one model up against its own wording. TypeSafe’s jaggedness notes already say two phrasings need not agree. Here the share that moves by 0.20 or more is 45%.