Which problems does Jev send to the big model?
Qwen3-4B answers all 164 HumanEval problems first. Jev reads each answer and scores how likely it is to be correct. Answers scoring below the threshold are discarded and the problem goes to Qwen3-32B. Each dot below is one problem, placed by its score. Drag the threshold and watch where every problem ends up.
Tests
When 32B returns a placeholder
Every threshold from 0.02 to 0.98. The ringed point is the current setting.