# POC scope and source

Source: [[POC] 데이터 전처리 방법론 검토](https://onepredict.atlassian.net/wiki/spaces/TD1/pages/3275522074/POC). Read from the authenticated Confluence browser on 2026-09-14; page displayed “Updated Sep 07.” The MCP connector still required reauthentication, so the browser was used for read-only access. No Confluence content was changed.

The page's objective is location-wise forming-defect risk using CAD/mesh or measured point clouds, material/product information and process conditions. Candidate intermediate targets are thermal histories, cooling rate, temperature gradient, pressure, deformation and thickness fields. Its proposed stages are initial point/image localization, geometry/thermal graph modeling, then constrained condition optimization.

| Representation requested by POC | Local preparation entry point | Model family |
|---|---|---|
| Multi-view images and depth/physical maps | `representations.py` | 2D CNN / U-Net |
| 3D voxel occupancy and fields | `representations.py` | 3D U-Net / sparse alternatives |
| Point cloud + per-point feature vectors | `representations.py` | PointNet / DGCNN |
| CAD/CAE mesh topology with global process vector | `cad_graphs.py` | GNN / MeshGraphNet |
| B-Rep face UV grids and adjacency | `brep.py` | UV-Net-style face encoder and GNN |
| Regular geometry/thermal field + conditions | Reuse multi-channel grids; actual thermal fields missing | FNO / neural operator |
| Process and geometry summary vectors | `representations.py` | Tabular surrogate + constrained Bayesian optimization |

Architecture names identify compact research baselines unless the model README explicitly states otherwise. Executing a family does not reproduce a cited paper's dataset, training budget or reported performance. See the generated model report for the implemented, executed, skipped and unavailable statuses.

The current real inputs permit geometry preprocessing and unlabeled model execution. They do not permit real supervised comparison of defect models, thermal surrogates or optimal forming conditions. No fields or labels should be invented to hide that gap. Synthetic examples are separate and explicitly named.

Implementation was delegated through the requested CLI configuration:

```sh
codex exec -m gpt-5.6-terra -c 'model_reasoning_effort="high"' ...
```

The three independent scopes and their invocation logs are stored in `results/delegation/`. The parent agent integrates, reviews and verifies the work.