ProVer fixes GRPO's credit assignment by verifying only pivotal decisions

Fixing GRPO's credit assignment problem without evaluating every step

GRPO assigns the same trajectory-level advantage to every token, blurring which decisions actually mattered. ProVer uses an agentic judge to contrast successful and failed rollouts and propose a segment that may explain the difference. Instead of trusting the judge, it verifies the segment by comparing terminal success rates of continuations sampled before and after it, then folds positive estimates into GRPO advantages. On ALFWorld, WebShop, and SearchQA, ProVer beats GRPO by 9.91% and 7.12% for Qwen3.5-2B and Qwen3.5-4B.