I built IMGNet – a face verification model that identifies people using sign patterns, not cosine similarity [R]

I want to share something I've been building as an independent researcher from Indonesia. TL;DR: Face verification model that replaces cosine similarity with sliding window sign pattern matching. Achieves 96.27% on LFW (pre-aligned) with a 10.58 MB model trained on CASIA-WebFace (490k images). When applied to ArcFace embeddings without retraining, IMG Sign Score gets 99.58% on LFW — only 0.24% below ArcFace+Cosine. The Motivation In Javanese, gratitude is "matur suwun" . In Sundanese, the same feeling is "h