ASSAY: RECOVERING FILTERED ENVIRONMENTS WITH VERIFIER-GROUNDED ACTOR–CRITIC RL

Code-derived reinforcement-learning environments are often selected using a screening model's rollout outcomes. Such selection mixes two questions: whether an environment is reliable, and whether a particular training recipe can use it effectively. We study this distinction in , where an all-pass/all-fail screen removes $2,628$ of $8,173$ tasks that have passed execution-consistency, leakage, and verifier-agreement checks. A unanimous training group gives GRPO zero reward-driven advantage, but a unanimous screening group does not establish that future training groups will be degenerate. We introduce , which re-admits these tasks, uses single-rollout asynchronous actor–critic optimization, and pretrains the critic on outcome-labeled records retained during environment construction. On the retained RL task pool, improves MiMo-V2.5 from GRPO's $54.4$ to $56.1$ on SWE-bench Pro and from $21.7$ to $24.3$ on DeepSWE. Expanding the pool raises these scores to $57.9$ and $27.5$. The same expansion changes  Val by $+2.7$ percentage points under and $-0.4$ under the evaluated GRPO configuration. Results on GLM-5.3-Flash show the same direction of improvement. Ablations, prompt-exposure controls, and training-cost measurements characterize the roles of value initialization, sampling, and asynchronous execution. These results support a practical conclusion: rollout screening should not be treated as an optimizer-independent verdict on training-data value.