Seal and Standard — physical-board recognition pilot

Recognition is performed locally in the browser. Models read piece placement;
side to move, rights and history come from the confirmed starting position.
Model confidence is not measured accuracy. This pilot needs independent full-game
evaluation, especially for unfamiliar wooden or custom historical pieces.

Models: cstr/chess-board-photo-onnx, MIT-labeled by their publisher.
Pinned commit: 4690fd5418ff7c405a76d1c156077ecd865ac053
https://huggingface.co/cstr/chess-board-photo-onnx
Pinned model card:
https://huggingface.co/cstr/chess-board-photo-onnx/raw/4690fd5418ff7c405a76d1c156077ecd865ac053/README.md
Occupancy SHA256: dcafa86708036f1c50d2471bcfef838a3973e63bc6e3eef08dd3adccf6e0067d
Piece SHA256: 79f7833e4fd802eb10de4d2765bb243ec51c40bb5097989af9ff4db789f2a8af

Crop implementation adapted from CrispChess and chesscog (MIT).
CrispChess copyright 2024–2026 CrispStrobe. Pinned source commit:
f029a319e35ce018936fc620dbf35a656a409b84
https://github.com/CrispStrobe/CrispChess
See CRISP-LICENSE.txt, CRISP-NOTICE.txt and CHESSCOG-LICENSE.txt alongside this file.
chesscog: Georg Wölflein and Ognjen Arandjelović, Determining Chess Game State
from an Image, Journal of Imaging 2021. https://github.com/georg-wolflein/chesscog

Model training provenance, as documented by the publisher:
- torchvision MobileNetV3-Small code: BSD-3-Clause. The publisher documents
  initialization from ImageNet-pretrained weights. The code license alone does
  not establish the terms for those pretrained weights.
- chesscog synthetic renders (OSF xf3ka), CC BY 4.0; Georg Wölflein and
  Ognjen Arandjelović. https://osf.io/xf3ka/
- samryan18/chess-dataset, MIT; copyright 2019 Samuel Ryan, Mukund
  Venkateswaran, Kurt Convey, Michael Deng.
- Roboflow100 chess-pieces-mjzgj, CC BY 4.0; original dataset creators
  Joseph Nelson and Brad Dwyer.
  https://universe.roboflow.com/roboflow-100/chess-pieces-mjzgj
CC BY 4.0: https://creativecommons.org/licenses/by/4.0/

The model publisher labels the exported ONNX files MIT. Torchvision cautions
that pretrained weights may have terms derived from their training datasets.
ImageNet's database access agreement limits database use to noncommercial
research and education. The reviewed documents do not establish how that
agreement applies to these fine-tuned exported weights; independent commercial
clearance for the inherited initialization has not been established here.
This provenance question does not relabel the ONNX files as noncommercial.
https://docs.pytorch.org/vision/stable/models.html#general-information-on-pre-trained-weights
https://image-net.org/accessagreement

Runtime: ONNX Runtime Web 1.30.0, Microsoft, MIT.
https://github.com/microsoft/onnxruntime
Runtime JS and WASM were copied from this exact npm package version.
Complete upstream v1.30.0 third-party notices are preserved alongside this file:
ORT-THIRD-PARTY-NOTICES.txt. These cover multiple build configurations; not every
listed execution provider or dependency is present in this browser WASM build.
Original: https://raw.githubusercontent.com/microsoft/onnxruntime/v1.30.0/ThirdPartyNotices.txt

Photo diagnostic: resized original photographs from samryan18/chess-dataset,
commit 62a8f5b0a45a931c91fd11bf6e1fe0c9fa0d25d9. MIT; see DATA-LICENSE.txt.
Original filenames: 5k2-8-8-8-2R5-8-K7-8.JPG and
5k2-8-8-8-7R-8-K7-8.JPG. Photos are separately staged positions, not chronological
game footage. They are in a dataset used to train the models and therefore
cannot establish independent model accuracy.
https://github.com/samryan18/chess-dataset