实验定位 混合偏正 / 论文机制更清晰
A32 直接修复 A31 暴露的机制问题：surface router 不能把 neutral 当 accept。于是本轮把整 case 决策改成三路：有强正证据则 accept repair，有强冲突则 reject repair，证据不足则 defer 到 A24 calibrated route。
结果：A32 repair selected=4，false reject=3，false accept=0，F@5 Δ=0.0088。相比 A31，它把 false accept 2→0，同时 F@5 Δ 从 0.0068 提到 0.0088。
实验设计（Image-2 风格）
模块设计（Image-2 风格）
Three-way Rule
- reject repair if surface_conflict_ratio > 0.42
- reject repair if mean_surface_delta > 0.010
- accept repair if surface_improved_ratio >= 0.45 or mean_surface_delta <= -0.015
- accept repair if surface_improved_ratio >= 0.22, surface_conflict_ratio <= 0.34, and mean_surface_delta <= 0.004
- otherwise defer to A24 calibrated route instead of treating neutral evidence as acceptance
Design Principle
A32 的核心不是多一个阈值，而是把证据语义拆开：positive evidence 支持保留 repair，negative evidence 要拒绝 repair，insufficient evidence 不应该被强行解释为成功，而是回退到更保守的 calibrated route。
实验结果（表格）
| 策略 | 选择 repair 数 | false reject | false accept | defer | selected Chamfer Δ | selected F@5 Δ | 选择的 repair case |
|---|---|---|---|---|---|---|---|
| A24 calibrated | 3 | 4 | 0 | - | -0.00054 | 0.0067 | gso_000_input4, gso_002_input4, gso_009_input4 |
| A31 binary surface router | 9 | 0 | 2 | - | -0.00015 | 0.0068 | gso_000_input4, gso_001_input4, gso_002_input4, gso_003_input4, gso_004_input4, gso_005_input4, gso_006_input4, gso_007_input4, gso_009_input4 |
| A32 three-way surface router | 4 | 3 | 0 | 6 | -0.00040 | 0.0088 | gso_000_input4, gso_002_input4, gso_006_input4, gso_009_input4 |
A32 逐 case 三路决策
| case | decision | selected | surface improved | surface conflict | surface neutral | mean surface Δ | GT repair Chamfer Δ | GT repair F@5 Δ | FR | FA | 原因 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| gso_000_input4 | accept | repair_hybrid | 0.4801 | 0.1295 | 0.1351 | -0.02144 | -0.00160 | 0.0202 | ['strong positive surface evidence'] | ||
| gso_001_input4 | defer | baseline | 0.0294 | 0.0470 | 0.9236 | 0.00001 | -0.00105 | 0.0082 | yes | ['mostly neutral surface evidence; defer to A24 calibrated route'] | |
| gso_002_input4 | accept | repair_hybrid | 0.6640 | 0.1862 | 0.1379 | -0.02789 | -0.00377 | 0.0502 | ['strong positive surface evidence'] | ||
| gso_003_input4 | defer | baseline | 0.1505 | 0.1767 | 0.6727 | 0.00248 | -0.00005 | -0.0041 | yes | ['ambiguous surface evidence; defer to A24 calibrated route'] | |
| gso_004_input4 | defer | baseline | 0.0106 | 0.0194 | 0.9700 | 0.00036 | -0.00041 | 0.0032 | yes | ['mostly neutral surface evidence; defer to A24 calibrated route'] | |
| gso_005_input4 | defer | baseline | 0.0000 | 0.0000 | 1.0000 | 0.00006 | 0.00119 | -0.0079 | ['mostly neutral surface evidence; defer to A24 calibrated route'] | ||
| gso_006_input4 | accept | repair_hybrid | 0.2412 | 0.2970 | 0.2676 | 0.00274 | 0.00144 | 0.0204 | ['moderate improvement with bounded conflict'] | ||
| gso_007_input4 | defer | baseline | 0.0055 | 0.0275 | 0.9670 | 0.00101 | 0.00281 | -0.0188 | ['mostly neutral surface evidence; defer to A24 calibrated route'] | ||
| gso_008_input4 | reject | baseline | 0.1739 | 0.7091 | 0.1170 | 0.01540 | 0.00474 | -0.0168 | ['surface conflict is high'] | ||
| gso_009_input4 | defer | repair_hybrid | 0.0152 | 0.0225 | 0.9623 | -0.00004 | -0.00006 | -0.0029 | ['mostly neutral surface evidence; defer to A24 calibrated route'] |
可视化结果
实验结论
- A32 验证了 A31 的问题来源：gso_005/gso_007 不是因为 conflict 检测失败，而是因为 neutral-dominant case 被二分类路由错误接受。
- 三路决策能消掉 A31 的 false accept，同时保持比 A24/A31 更高的 F@5 回顾收益；这说明“证据不足时 defer”是必要机制。
- 但 A32 仍不是最终方法：它还有 false reject=3，主要是 gso_001/gso_003/gso_004 这类收益较小、surface evidence 不强的 case。
- 论文角度的关键创新点进一步清晰：多视角非正交 mesh 生成不应只做候选排序，而应建模 observable evidence state：accept / reject / defer 对应可见改善、可见冲突和不可判定区域。
- 下一步应从 case-level route 进入 local surface edit：同一个 mesh 内，improved surface 锁定，conflict surface 回退，neutral/unobserved surface 交给生成模型补全。
下一步想法
- A33：做局部 patch 级三路 surface router。不要整 mesh accept/reject，而是在 surface sample/cluster 层面决定保留 repair、回退 baseline 或 defer completion。
- 把 A32 decision 转成 training/guidance signal：positive patch 做 consistency lock，negative patch 做 residual penalty，defer patch 不施加强约束。
- 在 100-case Benchmark 上做无 GT 相关性验证：surface_improved/conflict/neutral 三类比例与 Chamfer、F@5、watertight、nonmanifold 的相关性。
- 加入渲染指标：检查 A32 selected mesh 是否在 20 个测试视角上提升 mask IoU/LPIPS，而不只看 surface metric。