Spaces:
Running
Running
| title: Fiber Checker | |
| emoji: 🧵 | |
| colorFrom: yellow | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 6.28.0 | |
| python_version: "3.12" | |
| app_file: app.py | |
| pinned: false | |
| license: apache-2.0 | |
| short_description: Are papyrus fibers visible on a traced scroll surface? | |
| tags: | |
| - vesuvius-challenge | |
| - herculaneum | |
| - ct | |
| - papyrus | |
| - segmentation | |
| - mcp-server | |
| # Fiber Checker | |
| Are papyrus fibers visible on a traced Herculaneum scroll surface? Pick any traced segment of the | |
| [Vesuvius Challenge open data](https://scrollprize.org/data), or upload your own tifxyz mesh, and the Space renders | |
| the CT along the surface patch by patch and gives each patch a **fiber evidence**: near 1 where long, thin, parallel | |
| strands show, near 0.5 where the render looks no more fibrous than scrambled texture. | |
| - **Free, no key**: the default judge is a small network that runs on this Space's CPU. | |
| - **Claude as a second opinion**, with your own Anthropic key, used for that run only. | |
| - **A fiber map** on the segment's own mesh grid (`fiber.tif`, the same shape as `x.tif`), a one-page `report.html` | |
| for prize entries, the clearest and weakest patches, and a self-check against built controls on every run. | |
| - **Fiber or not?**: a game against the free judge, on renders of public scrolls on no prize list. | |
| - **Compare** two traces, patch by patch. | |
| - **API and MCP server** for scripts and agents. | |
| The Progress Prize asks entrants to "show visually that papyrus fibers are visible on your output surface" | |
| ([scrollprize.org/prizes](https://scrollprize.org/prizes)): this is a way to measure that. | |
| ## How the free judge was tested | |
| It was preregistered, then tested once on two public scrolls it never saw (PHerc0500P2, 120 patches; PHerc0343P, 62 | |
| patches; 2.2 um scans at level 2). It told the sheet from the gap beside it on every patch (pixel measure 0.79 and | |
| 0.76), and the sheet from a 14-pixel tile shuffle, a size never trained on, on 98% and 92% of patches (pixel measure | |
| 0.53 and 0.57). A first round failed on the tile shuffle and is kept on record. Jump hints, fiber direction and best | |
| depth are untested ideas. A clear map says fibers show on and near the surface (each patch is rendered at the | |
| brightest depth within 5 voxels); it does not prove the trace stayed on one sheet. Everything it shows is model output about texture: it never reads or finds text. | |
| ## Run it yourself | |
| The command line tool is `scripts/fibercheck/fibercheck.py` in | |
| [ChaseHendrick/Scrolls](https://github.com/ChaseHendrick/Scrolls); the weights are this Space's | |
| `weights/fibercheck-local.pt`. See the How-to guides tab. | |
| Scans and meshes: Vesuvius Challenge open data, CC BY-NC 4.0, read live from its bucket and never copied here. Code and | |
| weights: Apache 2.0. Made by Chase Hendrick. | |