Jev 五项公开测试拆穿 200 倍性能宣称
作者用五个可复现测试（自动研究、并行提问、技能建议、自动化框架、JevSQL）实测 Jev，把 193 倍宣称修正为 3.6 倍。
JEV DROPPED 5 PUBLIC TESTS THAT KILL A FAKE 200X CLAIM IN 4 MINUTES 01 AUTORESEARCH COOKBOOK 2,000 wine reviews became 67 numeric columns CatBoost moved from 3.09 RMSE to 1.77 02 PARALLEL QUESTIONS 13 decisions shared one 53,777-character document one batch cost $0.000497 versus $0.006090 across separate calls 03 SKILL SUGGESTION 488 requests against 182 skills wrong loads fell from 16.8% to 7.3% seven previously correct requests still broke 04 AY AUTOMATE HARNESS 791 labeled decisions the measured speed gap was 3.6×, not 193.6× 05 JEVSQL Choice, Score and Noul became real SQLite functions semantic judgment moved inside ordinary queries the stack worth stealing: retrieve with code → let Jev answer bounded questions → inspect confidence → escalate uncertainty → log corrections → keep irreversible actions behind policy the interesting Jev story is not that it wins every benchmark it is that the interface is small enough to test, replace and put behind deterministic safeguards marketing wants one giant multiplier builders need five public artifacts they can actually inspect all five sources, numbers and limitations are linked inside the article
这条是引用着另一条写的：x.com/fluixoo/status/2103123493567025321