{
  "author": "Codex",
  "role": "AI reviewer",
  "updated_utc": "2026-10-27T02:10:57.560604+00:00",
  "presentation_rule": "showcase-public-site/README.md and GeoBC-show/README.md",
  "reviews": {
    "R05": "核对最终验收记录：1000个有效结果由985条原结果和15个用户授权新工况组成，原15个未收敛ID排除且保留。十个原载荷目标、身份、有限性、正J、材料域与平衡验收记录完整。平衡验收不包含逐例稳定性、解的唯一性或实验符合性。首屏表面为原二次边界面分片；位移为真实倍数，节点应力仅作显示投影。",
    "R06": "核对冻结的嵌套TRAIN子集、原节点映射与输入/输出定义。56维输入均来自已知几何、边界、节点力及材料工况，没有真值响应或样本编号；归一化只由各自TRAIN子集得到。VM标签是原四GP单元最大值经参考体积加权的节点投影，不能表述为直接预测原GP场。该接口完成不构成验证/测试集性能结论。新增逐例工况说明来自冻结载荷与相位、转速、净压及厚度偏压参数。场景是理想化解释，可能响应是机理预期，不能据此判定失稳、损伤、疲劳或实际故障；具体几何和约束同样影响场响应。",
    "R07": "三组均完成固定2000轮；已核对曲线、保存预测、原节点指标和推理权重。20条实验位移平均/最差L2为14.875%/51.994%，节点VM为20.883%/35.960%。这说明本次预算下多工况响应尚未被充分拟合，不能只展示单例误差而称所有工况均已拟合。网络输出没有施加应力非负约束，逐例记录保留了负预测值，显示未通过裁零隐藏。未使用误差通过线，也未评估泛化。",
    "R08": "三组均完成固定2000轮；已核对曲线、保存预测、原节点指标和推理权重。20条实验位移平均/最差L2为2.117%/4.089%，节点VM为11.238%/15.898%。本次三档训练集上的平均误差均低于TS3；架构、参数量和每轮计算成本不同，结论限于这组配置与预算。应力仍是较大的误差来源；未以固定预算完成替代精度或泛化认证。"
  },
  "display": {
    "native_surface": true,
    "mesh_smoothing": false,
    "deformation": 1,
    "stress": "Original four-GP element-maximum VM projected to original nodes by reference-volume weighting.",
    "error": "U vector Euclidean difference or absolute projected node VM difference against saved FE targets; error on reference configuration.",
    "training_sample_selection": "Default: maximum squared relative U plus squared relative VM error within selected TRAIN subset.",
    "no_invented_accuracy_gate": true,
    "loads": {
      "source": "assets/rod-load-glyphs-v1/manifest.json",
      "geometry": "Same FESpatialGlyphs closed 3D arrows as C04; x-ray visibility for internal load locations.",
      "directions": "Frozen actual applied pressure/body-force directions, not displacement or reaction directions.",
      "motion": "Anchors follow FE/prediction displacement separately; error uses reference anchors.",
      "length": "Schematic direction samples, not proportional magnitudes."
    }
  }
}