拓展阅读 / JevBench：独立评测基准

知识库收录材料 · 原始文件 ↗ · 来源、版本与许可

Certo measurement — mapping fixed before the run

Written 21 Sep 2026 before the full run. altslate/certo-decision-model revision 02f596197d708d1c7779e3688dd6c976fd11afeb is the public v1 checkpoint: a ModernBERT-large encoder with a per-option scoring head, MIT licensed, run through the author's DecisionModel at Certo commit ed4d1333e5269418c1a519cd966a36b6559c0f15. The model file is 1,604,387,619 bytes, SHA-256 92735ee2a228cfeee4f90d322e42be698a209fcfd4d9d2d020252cfa02485789.

The frozen v1.2 set and scorer are unchanged. Certo exposes state + options, with no separate question field, so each JevBench item maps as follows:

• state: <question instructions>\n\n<state>;

• choice: one option per exact JevBench label, using its criterion as description;

• noul: no uses the false criterion and yes uses the true criterion;

• score: one option per exact level index, using the level text as description.

The checkpoint's own defaults remain fixed: state is truncated to 64 tokens and each option to 48 tokens. The complete native softmax is scored. The model runs serially after warm loading; no batch changes or retry-based answer selection. Endpoint condition and measured hardware are recorded in the row. Cost will use the same large-encoder hosted-price estimate as comparable ~400M encoder rows, times the model's retained input tokens; local or rented execution is never an automatic Cost score of 100.