tl;dr: TypeSafe's Jev model, a general-purpose classifier built on LLM token probabilities, has been rapidly adopted, but the author argues OpenAI could easily replicate it since they've been using tokens as micro-classifiers (for tool calls, stop conditions, etc.) for years. More importantly, OpenAI could fold classification directly into their frontier models via a `<prediction>` syntax, enabling calibrated self-checks during reasoning without leaving the GPU—useful for routing, safety, and stopping conditions. TypeSafe's only real moat is likely their synthetic training data and RL process; if that's shallow, OpenAI will out-build them quickly.