扩大训练数据与逐项模块

固定源选单模型在受控新城市达到安全门槛；真实20栋仍显示屋顶支持点删除风险，需要真实配对和独立核查。

• 源预选graph_20000_long_s20261031在新城市同污染机制下满足逐栋clean保护、失败与覆盖门槛。
• 同源训练teacher与模拟点云不能证明美国或真实AHN点云精度，也未证明标注节省。
• 附加真实AHN5探针的本栋足迹内class6仍有3492/37199记录被删除；它是生产分类风险诊断，非独立误删率。

完整汇总表

Family行是分别重建后seed损失均值；源选部署单模型另列。不能与概率均值集成或其他轮的绝对指标混同。

全部已生成主图、条件图与案例图

点图可打开原尺寸；阶段图保留原阶段语境，正式结论以本轮已审报告为准。

结果与逐栋下载

压缩文件为gzip JSON/CSV。科学数值、原始科学哈希和图像数据保持原值；公开复制品只改写本机路径和网页链接。原文件SHA与发布SHA分别列于出版清单，不能把公开复制品称为原字节文件。

适用范围与限制

• Automatic3DBAG Roofer-lineage teachers, not independent human GT
• Synthetic point observations, no real AHN/3DEP training claim
• Exact teacher roof-projection footprint and teacher ground are common priors; registration and footprint errors are not tested
• No optical CNN or spatial image fusion tested in this ladder
• No US precision, annotation saving, new architecture or SOTA claim
• Only screened modules were repeated; screening50 is part of source200, so module selection is exploratory.
• Three-seed family means average separately reconstructed errors, not the probability-mean ensemble.
• Source/final use identical12 simulated corruption mechanisms; new cities do not establish real-point-cloud accuracy.
• Native LoD2 status and edge-incidence diagnostics do not certify watertightness or complete solid validity.
• Terrain input changed before this round; old-round absolute scores cannot be subtracted to attribute model improvement.
• Bootstrap intervals omit teacher bias, new-city uncertainty, multiple-comparison selection effects and population seed uncertainty.
• 真实20栋来自source训练城市的固定ID，teacher与AHN同源；三partition均值是新的真实推理适配，分数未作真实域校准。
• RGB2023只用于质量预览，原prepared optical_features8仍缺失，主实验没有RGB输入。