Bundle contents

Supporting material for SUBMISSION.md — the cross-run GrowPatch sheet-switch detection writeup.

• SUBMISSION.md — the submission itself.
• FINDINGS.md — full research log (scanning methodology development, intermediate results, all 28-range sweep, CT ground-truth work) that SUBMISSION.md is distilled from. Read this for the complete reasoning trail, including dead ends investigated and ruled out along the way.
• scripts/ — every script used to produce the results in both documents, in the order they were developed:
  • cross_run_consistency.py, batch_scan.py — original coarse (45.5um) cross-run prototype and full-corpus scan.
  • fine_scan.py, visualize_fine.py, fine_scan2.py — fine (2.4um) follow-up and coverage-gap vs. switch classification.
  • characterize_switch_w010.py — first per-cluster normal/tangential displacement decomposition (superseded by switch_scan_v2.py for scanning, kept for the original analysis).
  • switch_scan_v2.py, region_merge.py — the corrected deep scanner (walks the full ranked disagreement list, not just top-N) and region re-clustering used for all final results.
  • sweep_remaining.py, rollup.py — automated the remaining 21-range sweep and produced the final 28-range summary table.
  • find_peak.py — extracts true peak-disagreement points per region (not bounding-box centers) for accurate spot-checks.
  • render_w107_patch.py — local tangent/normal-plane point-cloud visualization (pre-CT visual confirmation).
  • ground_truth_ct.py, ground_truth_ct_generic.py, boundary_check.py — raw CT ground-truth confirmation: streams CT chunks from the open-data volumes bucket and overlays both runs' points on the real material-density background.
  • prevalence_check.py — supporting data-volume sanity check for the "clean" (zero-signal) ranges.
• figures/ — the key result images referenced in SUBMISSION.md:
  • sideview_w107-109.png — pre-CT visual confirmation (two coherent parallel point-cloud bands).
  • ct_groundtruth_w107-109.png, ct_groundtruth_w073-076_tight.png, ct_groundtruth_w010-027_peak.png, ct_groundtruth_w064-068_wide.png, ct_groundtruth_w101-103_wide.png, ct_groundtruth_w104-106_wide.png — raw CT overlay confirmation at the 6 spot-checked peaks.
  • boundary_check_w107-109.png — before/inside/after divergence check ruling out a constant registration offset.

Reproducing

All scripts are pure Python (numpy, scipy, Pillow, matplotlib, stdlib urllib/concurrent.futures for HTTP chunk streaming) — no C++ build, no volume-cartographer install. They read directly from the public, CC-BY-NC Vesuvius Challenge open-data S3 bucket:

• s3://vesuvius-challenge-open-data/PHercParis4/segments/ — the two independent GrowPatch tracings (2026-06-23 and 2026-07-01 batches) used for cross-run comparison.
• s3://vesuvius-challenge-open-data/PHercParis4/volumes/20260411134726-2.400um-0.2m-78keV-masked.zarr — the raw CT reference volume used for ground-truth confirmation.

Scripts expect a local fine2_4/<run>_<range>/ mesh cache populated by sweep_remaining.py (or the original coarse cache via batch_scan.py); see each script's top-of-file VOLUME_URL/path constants for the exact layout.