The article argues that relying on models to draft forms leads to errors because they cannot distinguish between missing information and invented values, causing users to approve unchecked fields. It proposes shifting the goal from execution to generating a list of unknowns by using an external checklist that dictates what questions to ask.
- The system uses three checklists: fixed (tool choice, conditions), provider-specific (required fields, state changes), and user-specific (intent, context).
- A defined search order determines slot sources: user answer, instruction, preset values, observed values, or prior state.
- The approach treats "unknown" as a legitimate state after all sources are exhausted, rather than allowing the model to infer or fill blanks.
- It distinguishes between "no result" from a lookup and "unknown" remaining after the search ends.
This method prevents hallucinated data by forcing explicit confirmation of missing values through external constraints rather than model inference.