Jev 输出不满足互补性，正反查询置信度都高
作者实测发现 Jev 的 Choice/Score 输出里，一个命题和它的否定加起来不等于 1，二元结果两边都能拿到高置信度。
Within three weeks of Jev shipping on September 15, fourteen competing AI labs rolled out clones with prices already pushed to $0. Cloudflare, Perplexity, OpenAI, Liquid, Upstage, Together, and Inception all matched the exact same output schema: Choice, Score, Noul. The issue is what came along with that schema. Jev's own documentation notes that its outputs fail basic complementation. A query and its direct negation do not sum to 1. You can poll the endpoint on both sides of a binary outcome and receive high confidence on each, with no built-in state for uncertainty. A cheaper token will not fix that failure mode. An expanded context window will not either. If you are routing production logic through fast decision models, test whether complementary prompts add up to one before relying on the output. Zero-dollar pricing means nothing when the underlying math contradicts itself.