unaveraged Copyright 2026 unaveraged contributors Original project code, build tools, tests, documentation and icons are provided under Apache License 2.0. See LICENSE. The adapted JavaScript identified below retains its MIT license. Each browser package contains only its selected model and adapter; third-party licenses remain intact. 1. Tiny Receipt Reader Digit Slots 2M Source: https://huggingface.co/ivere27/tiny-receipt-reader-digit-slots-2m Revision: c9fa07886f99334ea820bfe312d6b4771d9460b0 License: Apache-2.0; model/LICENSE.txt in the digit package, downloaded from the pinned public source release. Changes: FP32 ONNX conversion through VolvoxAI. The default conv4-linear2 profile uses joint weight QAT, INT4 convolution weights and INT2/group32 dense weights, followed by static INT8 convolution calibration. Packed dense layers use dynamic INT8 arithmetic; other selected operations remain FP32. The optional W8A8 build calibrates INT8 convolutions. Exact graph/weight hashes and fitting algorithms are recorded in build-info.json and model/manifest.json in each corresponding browser package. Default builds download the public FP32 source, train and calibrate with synthetic receipts 000000 through 000223, and reserve 000224 through 000255 for validation. Weight training uses CPU by default, with CUDA available through QAT_BACKEND; convolution calibration uses CPU. Historical evaluated artifact identities and scope remain in digit/frozen-profiles.json and benchmarks/ in source for optional explicit imports. Fresh builds can produce different weights. 2. Tiny unaveraged VQA Structured QA 21M Source: https://huggingface.co/ivere27/tiny-receipt-vqa-structured-qa-21m Revision: f536c6ed181ca88929214f0bbc08f126f148c7ae License: Apache-2.0; original model/LICENSE.txt and model/NOTICE.txt in VQA ZIP, downloaded from the pinned public source release. Changes: FP32 encoder/decoder conversion through VolvoxAI; masked question padding to 192, batch 1, int32 IDs/masks, positive-length masked KV sentinel, SiLU numerical migration and common subexpression elimination. The default W4A8 PTQ profile packs 67 dense layers at INT4/group32 with dynamic INT8 activations; 13 convolutions use static INT8. Identical constants and token embeddings share storage. Other selected operations remain FP32. Default builds download the public FP32 source and perform PTQ with all 256 bundled synthetic receipts. Generated graphs, weight shards and manifests are stored in local caches before packaging. Historical evaluated artifact identities remain in vqa/frozen-profiles.json for optional explicit imports; fresh builds can differ. Model provenance and modifications are recorded in each package's build-info.json and model/manifest.json. Private calibration/evaluation images and questions are not redistributed with the public source or either browser package. The public documentation includes two unchanged synthetic Korean receipt examples from the pinned model repositories, together with selected public example questions and expected answers. These are not captured user receipts. Their original URLs, revisions and SHA256 hashes are listed in docs/assets/examples/sources.json. The Apache-2.0 license is in the root LICENSE. The original VQA notice is retained below. Documentation examples are not added to the extension packages. The public source also includes synth/: 256 original synthetic receipt PNGs and their generated content/glyph-coordinate annotations, distributed under this repository's Apache-2.0 license. Their image pixels are procedurally rendered; store names, address prefixes and item names come from a text bank built from existing receipt annotations. Captured receipt images and their original annotation files are not included. Generation uses rendered Noto Sans CJK and DejaVu glyphs; no font binaries are distributed in this dataset. The dataset manifest records image/annotation hashes and generation provenance. Calibration data is included in source exports and excluded from extension packages. 3. VolvoxAI 0.9.0 lite JavaScript/WebAssembly runtime Source: https://github.com/ivere27/volvoxai Distribution: published npm volvoxai-lite@0.9.0. Exact source and binary hashes are recorded in build-info.json. Copyright (c) 2026 VolvoxAI Contributors License: MIT; full text below and vendor/LICENSE.txt in each extension. The selected runtime files are included without modification. 4. Adapted model-specific JavaScript digit/adapter.js question routing: VolvoxAI examples/receipt_digit_reader/questionRouter.js. vqa/tokenizer.js: VolvoxAI examples/tiny_receipt_vqa/TinyReceiptByteFallbackBPEVocab.js. Copyright (c) 2026 VolvoxAI Contributors License: MIT; full text below, included in NOTICE.txt in each ZIP. These components retain their MIT license and attribution. Build/test dependencies retain their own licenses and are installed separately through npm and PyPI. Versions are recorded in package-lock.json and requirements.txt. ONNX Runtime is only the independent validation reference, not the PTQ engine or extension runtime. Build/test dependency trees are not redistributed. NAVER and Chrome are trademarks of their respective owners. These independent extensions are not affiliated with or endorsed by NAVER or Google. Original VQA attribution notice ------------------------------- TinyReceiptVQA Structured QA Copyright 2026 ivere27 This model package is distributed under the Apache License, Version 2.0. VolvoxAI MIT license (runtime and adapted JavaScript above) ---------------------------------------------------------- MIT License Copyright (c) 2026 VolvoxAI Contributors Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions: The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.