Claude Hand-Writes Neural Net Weights to Detect Fake Legal Citations, Beating Trained Models

Users had Claude hand-write all weights of a neural network for detecting fake legal citations without any training. It reached 97.8% accuracy on the LePhantomCite benchmark, beating trained models at 94.6%, and runs at 1.8M citations per second on Apple Neural Engine.

2026-11-04 ~ 2026-11-05 · 2 related posts

• Claude hand-writes a neural net's weights to catch fake legal citations, beating trained models — 141_1337 · 2026-11-04
• Claude hand-writes a legal-citation detector net, 1.8M cites/s on Apple ANE — paul_cal · 2026-11-05