Technical deep-dives on how Kyokon works: why we count recipe usage instead of asking an LLM, how canonical naming works, and the architecture decisions behind a nutrition API you can actually audit.
A deterministic scoring algorithm for matching recipe ingredients to USDA FDC foods. Five signals, zero machine learning, and the invariant that 'oil' must never match 'boiled.'
Why 'probability' is a euphemism for 'unmanaged risk' in safety-critical systems. A pattern language for building auditable ontologies from usage data instead of LLM inference.
Instead of using machine learning to infer what 'ground beef' means, we count how many times real recipe authors wrote 'ground beef' and use that as the canonical form. The wisdom of crowds replaces the wisdom of weights.