Using Jev's own confidence to decide which answers need a second look
An outside benchmark of 1,342 five-option rule-following questions ranks Jev 1.14 tenth of 26 decision models (1,030 right by the author's private key; Mammoth-1 1,139, Drex Max 1,127). Without the key I couldn't score answers, so I ran all 1,342 through Jev 1.14 and Drex Max and looked at where they split. They agreed on 1,061 (79%). Jev's median confidence was 0.97 where they agreed and 0.26 where they didn't. Sending every Jev answer under 0.7 to a second model would have sent 41% of questions on and caught 94% of the disagreements (at 0.5: 31% sent, 82% caught). On step-by-step procedure questions they agreed on only 12 of 48 and Jev's median confidence was 0.07. Jev: 140 ms median, $0.033 for all 1,342; Drex Max 271 ms, $0.028.
What I learned
Jev's confidence is a usable escalation signal; cascade low-confidence answers to a bigger model instead of trusting every choice. Set the cutoff on your own labeled rows. Keep procedures, counting and swaps off a decision model alone.
status: measured · questions: one five-option choice per question · result: Jev conf <0.7 flags 41% of questions and catches 94% of Jev/Drex Max disagreements · where it lives: System 1 Decision Benchmark, all 1,342 public questions