Same question, two inputs: Jev confidence jumps from 0.61 to 0.93
A hands-on test shows that replacing a conclusion-style state with separate evidence fields raised Jev's confidence on which worker runs next from 0.61 to 0.93.
i asked my agent the same question twice and got two different answers. the only thing i changed was how i wrote the input 0.61 became 0.93 the question: which worker runs next > vague state ("research_done": true): jev says write, p = 0.61 > separate evidence fields (sources_found, gaps, last_error): jev says research, p = 0.93 nothing else changed what actually got me is that the state matters more than the question. we rewrite prompts for hours and hand the model a conclusion instead of evidence the rule i use now > policy goes in the instructions and the option text > evidence goes in the state, not conclusions > pass last_error and the sources found, never "done": true > act on high confidence, review the middle, escalate the low illustrative example from a public test. confidence is not accuracy, so calibrate your floor on your own labeled examples save this before you build your next router what's the vaguest field in your agent's state right now? full stack in the article below ↓
This post quotes another post: x.com/kiruwaaaaaa/status/2109593981122195661