| """veil_pgd: image-side training data protection. | |
| Applies low-visibility perturbations to public images so vision-language | |
| training-data ingestion pipelines receive less reliable captions, tags, | |
| embeddings, or synthetic labels. | |
| The Mac orchestrator (this package) stays torch-free; all model inference runs | |
| on klaus-3 behind HTTP endpoints. Paid black-box validation runs on OpenRouter. | |
| """ | |
| __version__ = "0.1.0" | |