"We measured this. Not estimates — actual token counts using the cl100k_base tokenizer against real schemas, verified by an automated test suite."