A follow-up example from Microsoft’s ThinkingBox uses scores from checking a database’s actual state after an agent completes a task as reinforcement learning rewards. In an experiment training Qwen3.8-27B for 50 steps, single-attempt success rose from 51.70% to 60.78%, and the number of tasks solved successfully in all 20 attempts increased from 38 to 107. However, the number solved at least once in 20 attempts fell from 453 to 446. Unlike the earlier introduction of the evaluation tool, this report focuses on applying it to training. Consistency, average success and the breadth of solvable tasks need to be measured together to judge the improvement.