One Jev pass scores 1,072 translations for $0.3165

七誌 posted on September 23 that a multilingual translation test had been scored again, this time with one Jev pass instead of three. The earlier ruler was Qwen 3.6, and the median of three runs was there because a temperature-zero run froze one of the wobbles. The post says Jev moved by about an error bar, so one evaluation was enough: 7.5 million tokens, $0.3165, 5 minutes 23 seconds. The log is examples/tr/onde/JEV.md in 7shi/multilingual-reader. The repository license is CC0.

The run is dated September 22 in that note. The model is pinned to jev-1.13.0, not the jev-latest alias. The scheme id is degrees@f518286e. Each language gets five Score questions over the same five severity levels. The levels run from 0, the criterion fails across the whole document, to 4, it never fails. Each level is worth 5 points on a 0 to 20 scale, and five criteria sum to the 0 to 100 number in the tables. The criteria are readability, fluency, terminology, contextual adaptation, and information completeness. Output tokens are 5,561 for every model, because a Score reply has no prose. Input tokens move with the length of the translation, from 459,674 to 489,167 on a single model, 6.4% across the corpus.

Sixteen models by 67 languages is 1,072 evaluations. Input tokens sum to 7,536,547. Output tokens sum to 88,976. Wall time is 5 minutes 22.6 seconds. The per-language clocks inside the files sum to 313.6 seconds. The other 9.0 seconds is startup and the resume scan, sixteen times. Cost is $0.3165, which the note splits as $0.0198 and 20.2 seconds a model, and $0.000295 an evaluation. Median time per language is 0.28 seconds, with a minimum of 0.22 and a maximum of 2.21. An earlier estimate in the plan was about $0.21 and five minutes. The clock matched. The bill was 51% higher. The note says later quotes should use $0.32.

The reason one run replaced three is a prior check the note attributes to experiment 12: this evaluator’s run-to-run range was 1.12 points, and run 1 alone matched the median of three at Pearson 0.998. A port check rescored four models that an earlier script had already sent to the same pin and the same rubric. Input token counts matched model for model. Over 268 translations the mean signed difference is -0.006 points, Pearson 0.9995, Spearman 0.9980. Mean absolute difference is 0.517 and the median is 0.375. Twenty-six of the 268 differ by more than 1.12 points. Twelve came back bit-identical. The note says the evaluator was never claimed to be deterministic.

The published grid is in the translation README. Quality bands stay where they were under Qwen 3.6: 90 and above, 80 to 89, 60 to 79, and below 60. The note says Jev’s scale is narrower at the top, so fewer languages reach 90 without any change in the translations, and that the line at 80 is the one that still means what it meant. On the gemma4 core-language table, which is one translator and five languages, the averages are German 89.88, Japanese 87.00, French 86.93, Chinese 86.20, and Spanish 84.92. A separate reference-translation table, proofread by other models, sits higher: English 91.80, German 91.60, Spanish 90.70, Chinese 90.55, Japanese 87.85. That table is the proofread reference pages. The 16-model grid is the bake-off.

On the 16-model grid a few cells show how wide the same ruler gets. French: Ox Alpha 95.9, gemma4 91.1, GPT-OSS 78.3. Finnish: GPT-5.6 Luna 91.6, gemma4 90.6, Bonsai 2 27B 3.8. Latvian: Union Alpha 91.2, gemma4 40.9. The README keeps the full language list. We did not rescore a file.