Jake-vc's picture
Add files using upload-large-folder tool
e88a3a0 verified
|
Raw History Blame Contribute Delete
3.84 kB
---
license: cc-by-nc-4.0
pretty_name: "First Light: PHerc0175A meshes, ink maps and screening results"
size_categories:
- 1K<n<10K
---
# First Light: PHerc0175A
Surfaces, ink maps and screening results from a documented First Letters attempt on PHerc0175A (public ESRF scan
`20250521115057-8.640um-1.2m-116keV-masked`, 8.64 µm). The write-up, verdicts, readout rules and every script are in the
GitHub repository [jake-vc/first-light-pherc0175a](https://github.com/jake-vc/first-light-pherc0175a).
**In short:** two tall windows of the scroll were fitted with villa main's spiral fitter, every winding was snapped to
the papyrus and flattened. Window P was rendered and screened with three ink models; no ink was found, and the analysis
explains why PHerc. 343's ink thresholds do not transfer to this scroll. Window W2 is fitted and flattened, ready to
render, and was not screened.
## Start here
| If you want to... | Use |
|---|---|
| Render a winding yourself | `window_*/meshes/flattened/<winding>/` (tifxyz: `x.tif`, `y.tif`, `z.tif`, `meta.json`) with VC3D `vc_render_tifxyz` against the public volume above |
| Know how well a winding sits on the papyrus | `window_*/meshes/placement_snapped.tsv`: `frac_with_sheet` = share within ±2 voxels of the predicted sheet |
| Redo the ink analysis without a GPU | `window_P/ink_maps/<winding>/`: three models' maps at the surface, both sides |
| See what was judged | `window_P/judged_windings/<winding>/yellow_small.png` (reading side) and `yellow_blind_small.png` (control side) |
| Continue or re-export a fit | `window_*/fit/checkpoint/step_10000.ckpt` with villa main's `spiral-fitting` (commit in `fit/villa_commit.txt`) |
Windings are named `w<NNN>_main` (spiral turn number, inner to outer); the same number in both windows is the same turn.
## Layout
```
window_P/ slices 4,800-7,800, windings 10-129; screened (117 windings with >= 70% on the papyrus)
fit/ fit log, loss/placement curve, settings, villa commit, final checkpoint, export and check logs
meshes/
exported/ meshes as exported by the fit
snapped/ snapped onto the predicted papyrus sheet
flattened/ snapped and flattened: the surfaces that were rendered and screened
placement_exported.tsv, placement_snapped.tsv share of each winding on the sheet, before and after snapping
screen/ per winding: flag statistics (stats.json), contact sheet (sheet.png), render crop; summary.jsonl
ink_maps/ reader-v2, dense_native, reader-v3 at depth 0; *_reverse.tif = reading side, others = blind side
judged_windings/ the ten windings judged by eye: six-model whole-winding views (reading and blind side) and maps
window_W2/ slices 7,600-10,600 (200-slice overlap with P), windings 10-129; fitted and flattened only
fit/, meshes/ as above (113 windings with >= 70% on the papyrus)
analysis/
gate_ranking/ flag ranking and per-winding background statistics
rescreen_self_calibrated/ re-screen with thresholds set from the blind side; 343 control check
speed_benchmark/ screen timings and map comparisons
control_343/ PHerc. 343 positive control: maps and calibration
pipeline_logs/ job logs and state markers as run
MANIFEST.tsv sha256 of every file
```
Not included, because they can be re-created from public data: the CT, the community lasagna inputs (SharpMeow), and
the 66-layer renders.
## Licence and credits
CC BY-NC 4.0, as the data it derives from: the PHerc0175A scan and models (Vesuvius Challenge), the community lasagna
fields and center line (SharpMeow). Ink models: reader-v2/v3 (DomRusso2), dense_native, ink_9um and Hecate (Vesuvius
Challenge and the PHerc. 343 winning team). Built with an AI coding agent (Claude Opus 5.5, Claude Code).