Third-Party Notices

Models, code and data created by others, and the terms they carry

Version 2.13 · Effective 28 September 2026 · Alberta, Canada

Contents

• Summary
• 1. Bundled machine-learning models, ported code and word lists
• 2. Sample and demonstration data
• 3. Statistical methods and reference literature
• 4. Reporting a licensing problem

Summary

The Software bundles machine-learning models ported from Google’s MediaPipe project and from published research code, image-analysis code ported from open-source libraries, and the English word list from Bitcoin Improvement Proposal 39, under the Apache License 2.0 and the MIT License as noted below, together with Apple frameworks used under Apple’s own terms.

Each component is supplied by its authors on an “as is” basis, without warranties or conditions of any kind. Their authors have no involvement in OphthoRuler EyeStudio, do not endorse it, and take no responsibility for it.

ℹ️ Full model cards live in the app

The Model tab on macOS shows each installed model’s version, authors, licence, citation and source links, along with what it was trained on and the measurements it supports.

1. Bundled machine-learning models, ported code and word lists

• FaceMesh (facial landmarks): 468-point face mesh. Ported to PyTorch by George Grigorev from Google MediaPipe’s face_landmark model, then converted to Core ML. Apache License 2.0. Source: github.com/thepowerfuldeez/facemesh.pytorch; upstream Google MediaPipe, also Apache License 2.0.
• IrisLandmark (iris and eyelid refinement): Iris landmarks plus eye-contour and brow points. Ported to PyTorch by Cédric Ménard from Google MediaPipe’s iris_landmark model, then converted to Core ML. Apache License 2.0. Source: github.com/cedriclmenard/irislandmarks.pytorch; upstream Google MediaPipe, also Apache License 2.0.
• LightIrisNet (iris and pupil segmentation): Visible-light iris and pupil segmentation, MobileNetV3/DeepLabv3+. By Venkataswamy et al., released as part of an open-source framework for quality-assured smartphone visible-light iris recognition (arXiv:2512.15548). MIT License, Copyright (c) 2026 Naveenkumar GV. Source: github.com/naveengv7/LightIrisNet, which carries the licence.
• GazeTracking (pupil localisation for the motility chart): Classical image-processing pupil localisation used by the 9-gaze motility (extraocular movement) chart, ported natively to Swift from Antoine Lamé’s GazeTracking library; no Python or OpenCV code ships. MIT License, Copyright (c) 2019 Antoine Lamé. Source: github.com/antoinelame/GazeTracking, which carries the licence.
• BIP39 English word list (Emergency Recovery Kit): The 2,048-word English list used to spell the recovery code, embedded verbatim from Bitcoin Improvement Proposal 39 (bip-0039/english.txt in the bitcoin/bips repository), by Marek Palatinus, Pavol Rusnak, Aaron Voisine, and Sean Bowe. BIP 39 states that it falls under the MIT License; no separate copyright notice accompanies the list. Source: github.com/bitcoin/bips.
• Apple Vision, ARKit, and Core ML: Face-landmark detection used as a fallback landmarker, and the runtime for the models above. ARKit face tracking supplies the face mesh and depth used by the TrueDepth levator capture; Vision also locates faces for panel crops and reads the head’s angle off the camera preview for the capture screen’s en face check (frames are checked and discarded; nothing is kept). Supplied by Apple Inc. under Apple’s own terms as part of the operating system.

The Apache License 2.0 text is available at www.apache.org/licenses/LICENSE-2.0. It requires that attribution and the licence travel with the work, which this document provides, and it disclaims all warranties and limits the contributors’ liability.

The MIT License text is available at opensource.org/license/mit. It likewise requires that the copyright notice and the permission notice travel with the work, which this document provides, and it disclaims all warranties.

📌 Models are fixed and never learn from you

Bundled models are shipped as-is and are never trained, fine-tuned, or updated using your data. The Software contains no training capability.

2. Sample and demonstration data

ℹ️ No photograph of any person is bundled with the Software

The Software ships no photographs. Every photograph in your library is one you put there.

The Sample Studies folder contains synthetic, computer-generated data produced to demonstrate each study design. It holds no photographs and no measurement of any real person, and nothing in it may be reported as a research finding.

If your library contains a demonstration study titled “Sample - Ptosis 2026” (or “Ptosis Repair 2026”), its images are photographs of real people, and the measurements attached to them are simulated. Depending on which build created it, the images came either from the Chicago Face Database (supplied for non-commercial research use under terms that prohibit passing the images on to anyone who has not obtained them from the database) or from Unsplash, under a copyright licence granted by the photographer. Neither source conveys any consent from the person depicted and neither is a model release: do not export, redistribute, publish, or analyse those images, and delete the study if you would rather not hold it.

3. Statistical methods and reference literature

The statistical engine implements published methods and has been cross-checked against R and against canonical published datasets. Method citations appear in the in-app Statistics Guide and in generated manuscripts.

Neither the authors of those methods, nor the authors of the R packages used for cross-validation, nor the authors of any dataset used as a reference, are associated with the Software or responsible for it. Agreement on a test case is not a warranty that any result is correct.

4. Reporting a licensing problem

If you believe a component is misattributed, incorrectly licensed, or included without the right to do so, write to m1wo9uczq.hfrxpeablapu05r28ovnhkzt3uhf3p9aohu@5alj5au6gp2e2yl9x and it will be investigated and, where warranted, removed.

Text fingerprint (SHA-256): 1334bac67464e9f013789d39a744db455c1a3868590fdb6d8879d94138daeea4