Preprint tests label-free gate for optimization-modeling agents

A preprint introduces AdmitOR, a label-free admission gate for skills used by LLM-based optimization-modeling agents. The method resamples instances from an extracted parameter domain, compares value-function traces across model families and solver stacks, then returns accept, abstain or escalate using a calibrated threshold. On one collection of logs, the authors report admission precision of 0.927, versus 0.871 for majority vote and 0.726 for execution success. They also report that its preregistered false-discovery criterion did not hold on a wild stream.