ruleqa 7源同图多轮 p10 · 3bench 数据验证

训练 mj-mb-2026091117-36zg9q（qzy-sn6v0710-vst-ruleqa7ds-mt-p10-m3h20-4n-0911c，1000 iter 已完成）· 90% 固定 baseline + 10% ruleqa 候选数据 · no-think · 每个 step 均通过真实结果核验（20,108 条 prediction/eval 逐条核对、0 API 失败、题目内容跨 step 一致、deadline 7200s 确认）· 本页无 baseline 臂（collection 仅含本训练）· 同训练的 ERQA 曲线见 erqa.html

NaturalPerception Circular V3 · 准确率（%）

ruleqa 7ds mt p10（24.66% → 35.53%，峰值 step 1000 35.53%）
22.727.632.637.51002003004005006007008009001000NaturalPerception Circular V3 · step 100: 24.66%NaturalPerception Circular V3 · step 200: 31.14%NaturalPerception Circular V3 · step 300: 30.93%NaturalPerception Circular V3 · step 400: 32.08%NaturalPerception Circular V3 · step 500: 33.54%NaturalPerception Circular V3 · step 600: 33.75%NaturalPerception Circular V3 · step 700: 35.32%NaturalPerception Circular V3 · step 800: 34.69%NaturalPerception Circular V3 · step 900: 35.11%NaturalPerception Circular V3 · step 1000: 35.53%checkpoint

SolidGeo Perception Circular V2 · 准确率（%）

ruleqa 7ds mt p10（25.69% → 35.38%，峰值 step 800 36.09%）
23.828.533.238.01002003004005006007008009001000SolidGeo Perception Circular V2 · step 100: 25.69%SolidGeo Perception Circular V2 · step 200: 32.03%SolidGeo Perception Circular V2 · step 300: 33.91%SolidGeo Perception Circular V2 · step 400: 34.67%SolidGeo Perception Circular V2 · step 500: 34.82%SolidGeo Perception Circular V2 · step 600: 35.79%SolidGeo Perception Circular V2 · step 700: 35.74%SolidGeo Perception Circular V2 · step 800: 36.09%SolidGeo Perception Circular V2 · step 900: 35.53%SolidGeo Perception Circular V2 · step 1000: 35.38%checkpoint

Perception-geometry-mc Circular V1 · 准确率（%）

ruleqa 7ds mt p10（23.95% → 34.57%，峰值 step 900 35.29%）
21.927.032.237.31002003004005006007008009001000Perception-geometry-mc Circular V1 · step 100: 23.95%Perception-geometry-mc Circular V1 · step 200: 31.29%Perception-geometry-mc Circular V1 · step 300: 33.10%Perception-geometry-mc Circular V1 · step 400: 33.90%Perception-geometry-mc Circular V1 · step 500: 33.86%Perception-geometry-mc Circular V1 · step 600: 34.86%Perception-geometry-mc Circular V1 · step 700: 34.86%Perception-geometry-mc Circular V1 · step 800: 33.95%Perception-geometry-mc Circular V1 · step 900: 35.29%Perception-geometry-mc Circular V1 · step 1000: 34.57%checkpoint

数据核验时间：2026-09-12 UTC · 核验脚本 rule_qa_p10_m3h20_0911/monitoring/verify_threebench_snapshot.py