# Media and model attribution

## Recording and screenshot

`optiqc-demo.webm` and `desktop-result.png` were captured from the genuine local
OptiQC application on **October 3, 2026**. Model results in the recording were
computed by the application's real inference engine. There are no simulated
inference outputs. This publication is a static recording walkthrough; it does
not run a hosted inference server.

The visible images come from the **MVTec AD** dataset, created by **MVTec Software
GmbH**, under **Creative Commons Attribution-NonCommercial-ShareAlike 4.0**.
The recording and screenshot retain this license and attribution. The images
were not modified before inference; the application creates visualizations over
the images. Both artifacts reproduce the application interface and results.

- Original dataset: https://www.mvtec.com/research-teaching/datasets/mvtec-ad
- License: https://creativecommons.org/licenses/by-nc-sa/4.0/
- Included full license: `LICENSE-CC-BY-NC-SA-4.0.txt`
- Dataset mirror: https://huggingface.co/datasets/foersben/mvtec-ad
- Mirror revision: `c75b39616f84db43677bcc8228caaafaf5096d7f`
- Samples: `bottle/test/broken_large/000.png`, `bottle/test/good/000.png`,
  `bottle/test/contamination/000.png`

These examples demonstrate software behavior; they are not a benchmark and do
not establish accuracy or generalization. Dataset category labels are source
annotations, not predictions.

## Fitted model and implementation

The application integrates a fitted PyTorch feature-extractor checkpoint and a
FAISS nearest-neighbor index. It does not claim to have trained the pretrained
feature extractor. The original fitted checkpoint is an upstream component;
this release adds inference integration, the inspection interface,
visualizations, input validation, exports, and a limited image-threshold
calibration investigation.

- Application: https://github.com/sonybej-art/optiqc-anomaly-detection
- Fitted checkpoint: https://huggingface.co/javito3/optiqc-bottle-weights
- Pinned model revision: `43536dd1ee8bffb1b27b66b0058a38a2a3f6ca3b`
- Checkpoint SHA-256: `41823d6998eb31dd01bcf68aa3f94573594b02837aa06d3ea96dbfff8e19fb5c`

Anomaly scores are distances, not probabilities. Heatmaps indicate relative
anomaly intensity. Candidate hotspot boxes summarize heatmap regions and are
not object detections. The recorded threshold belongs to a limited normal-image
calibration pilot, not a validated production operating point.

## Page software

The static page software is licensed under MIT (`LICENSE`). That software
license does not replace the separate CC BY-NC-SA 4.0 license of the visible
MVTec-derived media.