A form did five things at once: judged the condition, picked which form to open, took the values, carried where each value came from, and validated the required fields.

Natural language kept the third one and gave the rest to the model. Nobody wrote down which ones went missing.

LLMs were trained by filling in blanks. Now that we've moved from conversation to action, we tell them not to fill in blanks.

But nobody has handed them a list of what they aren't allowed to infer, or shown them how to fill a blank without inferring.

Here's an attempt at that list and that method.

https://github.com/Jang-woo-AnnaSoft/execution-state-preflight/