# AC28 + ME1a preprocessing size decisions (expanded)

## What

Fifty AC28/ME1a candidate inputs were measured: images 64, 128, 256, 512, 1024, 2048; surface voxels 32, 64, 128, 256, 512, 1024; requested graph nodes 3k, 5k, 6k, 12k, 24k, 48k, 96k; CAD points 2048, 4096, 8192, 16384, 32768, 65536. Selected: six 512² images with 980 mm FOV; 64×8×40, 15 mm surface voxels; 5k requested CAD context graph; 8192 CAD points. The localized weak-risk output remains 12k.

## How

The first three sizes per modality come from `build.py`; extra image, voxel and point candidates come from `extend_sizes.py`; the three extra graph budgets are generated from each source STL with `전처리/cad_graphs.py`. Run `finalize_extended.py` after those steps. It uses a measured 2048 raster as the high-resolution image reference, one common physical camera per part for graph comparisons, and the existing occupied-center/CAD-node distance measures for voxel and point. Run `validate_standardized.py`, `render_candidate_samples.py`, and `publish_candidate_gallery.py` afterward. Raw values: `measurements.csv`, scores: `decision_scores.csv`, size contract: `standard_sizes.json`. The browser report is `index.html`; every candidate has a PNG in `candidates/`.

## Why

The smallest candidate passing its gate wins. The explanatory score is fidelity 50 + spatial detail 30 + relative input cost 20. Image needs 2048-reference six-view silhouette IoU ≥0.975 and ≤2 mm/pixel, corresponding to at least five pixels across the 10 mm weak-anchor scale. The worst 256 image IoU is 0.9526; 512 reaches 0.9788 at 1.918 mm/pixel. The old 256 self-reference IoU 1.0 was inadequate evidence. A full six-view 17-channel 512 float32 tensor takes 107 MB; 2048 takes 1.71 GB. The voxel gate is P95 occupied-center distance ≤15 mm and pitch ≤15 mm; graph context gate is surface P95 ≤0.35 mm and fixed-camera IoU ≥0.99 to the 12k graph; point gate is CAD-node coverage P95 ≤10 mm.

These are preprocessing decisions, not proof of trained-model gain. Historical CNN and voxel checkpoints used 128 image and 32 voxel inputs. The present 12k risk GNN uses 12k directly; 5k geometry context does not reproduce the same-run ME1a 12k risk score under its current nearest-node output mapping. ME1a weak labels and registration remain unvalidated.