ForgeSight-AD v0.1.0 - Public portfolio release

Highlights

• Evidence-first industrial anomaly detection portfolio covering all 15 MVTec AD categories.
• Audited comparison of PatchCore, DINOv2 nearest-neighbour features, PaDiM, and a fixed Train-only fusion rule.
• Responsive zero-install GitHub Pages case study plus a separate local FastAPI inference lab.
• Path-sanitized aggregate reports, publication manifest, experiment protocol, metric definitions, and explicit claim boundaries.
• Deterministic exact chunked kNN: equal-distance ties resolve to the lowest global reference index.

Measured outcomes

• Best image AUROC: 0.9728 (tail-gated PatchCore + DINOv2 fusion)
• Best pixel AUROC: 0.9747 (PaDiM)
• Best AUPRO@0.30: 0.9237 (PaDiM)
• Evaluation scope: 1,725 official Test images across 15 categories

Release quality

• Python 3.10, 3.11, and 3.12 contract tests on GitHub Actions
• Ubuntu and Windows public-snapshot and fresh-wheel checks
• GitHub Pages deployment verified
• Code-only wheel: forgesight_ad-0.1.0-py3-none-any.whl
• Wheel SHA-256: 8a2bfa292b5072b373821f1ffcc437102ea052e21ac6c602d26c801c6325b8b4

Public boundary

This sanitized release does not include MVTec AD images or masks, pretrained weights, memory banks, model bundles, logs, or predictions. The published benchmark reports preserve aggregate evidence from documented private clean runs; this public tag does not claim to reproduce those GPU benchmarks by itself.