STOCKFISH.IO IMAGE RECOGNITION — 21 OCTOBER 2026

Fenshot 0.1.5, source commit d10613b4dc720a8ed9716eab70ea681103e12692:
https://github.com/scoriiu/fenshot/tree/d10613b4dc720a8ed9716eab70ea681103e12692
Copyright 2026 SORTINO LABS S.R.L. (kibitzo.com), MIT.
The package ships chess-tiles-v2.onnx with its source and MIT licence. Its
README/training scripts describe weights trained from scratch on synthetic
boards made with Lichess/chess.com graphics. No separate restrictive weights
licence was found in the pinned package. The model is copied without changes;
training graphics are NOT distributed with this site. Retain the included MIT
licence. This records the upstream provenance, not an independent audit of all
upstream training-data permissions.

The pure detector/tiles/FEN functions are bundled in core.js with esbuild 0.25.12
(format esm, minify, bundle). Exports: findChessboardCorners, snapCorners,
extractTiles, rgbaToGray, probsToPlacement, resolveOrientation, flipPlacement.
Sources: packages/fenshot/src/{detector,tiles,fen}.ts at the commit above.
The upstream detector credits Elucidation/tensorflow_chessbot, MIT, copyright
2016 Sameer Ansari. Its exact notice is included beside the fenshot licence.
The library's automatic castling/FEN helpers are NOT used.

Model input: tiles, float32 [64,1024], 32x32 grayscale tiles in A1..H8 order.
Output: probs, [64,13]; classes empty,K,Q,R,B,N,P,k,q,r,b,n,p.
Scores below 0.70 flag uncertain squares. These are uncalibrated model scores,
not probabilities of a correct board. Pawn-rank orientation is only a hint.

ONNX Runtime Web 1.26.0, Microsoft Corporation, MIT:
https://github.com/microsoft/onnxruntime/tree/v1.26.0
https://www.npmjs.com/package/onnxruntime-web/v/1.26.0
Unmodified npm distribution: ort.wasm.min.mjs, ort-wasm-simd-threaded.mjs and
ort-wasm-simd-threaded.wasm. LICENSE.txt and ThirdPartyNotices.txt are from the
same v1.26.0 upstream tag. Runtime numThreads=1, proxy=false, WASM execution
provider, self-hosted versioned paths. The binary's name includes "threaded"
but this configuration uses one thread without SharedArrayBuffer/COOP/COEP.
Source maps are not required or included. No WebGPU path or remote fallback.

See ASSET-SHA256.txt for exact byte counts and SHA-256 hashes.
No browser-side model training, paid API, secret key or inference server.

Scope: conventional 2D screenshots and diagrams, not physical 3D pieces.
Flat perspective rectification is experimental. Source rotation, coordinate
orientation and main-board flipping are separate operations. Always review.