When to and when not to use return validators

The previous recommendation to always use return validators in the Convex backend platform, particularly when working with LLMs, is now being revised. While these validators enforce an exact data contract at runtime, their mandatory use creates verbose, token-inefficient code that is difficult for AI agents to process and increases the risk of model hallucinations. This practice also makes code harder to refactor, as schema changes require updating numerous duplicated validators. The new guidance suggests relying on TypeScript inference by default and reserving return validators for specific cases where runtime enforcement is crucial, like with static codegen or unvalidated external data.