https://stanleyngugi.netlify.app/ Technical writing by Stanley Ngugi on AI for formal methods — LLM theorem proving, grammar-constrained decoding, RL environments — and machine learning research. en-us https://stanleyngugi.netlify.app/posts/2026-08-22-grammar-constrained-decoding-lean.html https://stanleyngugi.netlify.app/posts/2026-08-22-grammar-constrained-decoding-lean.html Sat, 22 Aug 2026 00:00:00 +0000 Grammar-constrained decoding makes 100% of generated Lean tactics parseable at zero latency cost — while hallucination stays at ~50%. That asymmetry is the point. https://stanleyngugi.netlify.app/posts/2026-08-22-lean-tactic-language-cfg.html https://stanleyngugi.netlify.app/posts/2026-08-22-lean-tactic-language-cfg.html Sat, 22 Aug 2026 00:00:00 +0000 A 53-production context-free grammar for Lean 4 tactics covers 99.86% of every tactic line in Mathlib — and the methodology transfers to Rocq and Isabelle. https://stanleyngugi.netlify.app/posts/taming_polysemanticity.html https://stanleyngugi.netlify.app/posts/taming_polysemanticity.html Sun, 19 Oct 2025 00:00:00 +0000 An exploratory study using sparse autoencoders to measure how L1 vs L2 regularization, initialization, and activation choices shape polysemanticity in overcomplete networks.