Select, Don’t Compose

Ask a language model an open question and it will give you an answer. Ask it one it can’t answer and it will still give you an answer. That’s the whole problem with putting AI in charge of decisions that matter, and most of the industry’s response has been to ask the model, more and more politely, to be careful.

I tried that. It doesn’t hold. What worked was changing the job.

The wrong job

Language models are most likely to make things up when they’re asked an open question, when they’re writing at length, and when the fact that decides the answer is buried in noise. Most production prompts do all three. They’re long, polite and descriptive, and they end with “please respond with your assessment.”

A different job

Writing an answer and choosing one are different tasks. A model asked to write will invent something plausible when it doesn’t know. A model asked to choose either finds a match or it doesn’t, and “doesn’t” is an outcome a system can handle honestly.

That’s the idea behind PWTL, the task language I built for GARDE1. Every decision the platform makes is a choice, not a composition, and it arrives with its evidence and its reasoning attached. It’s patent pending, so I’ll stay at the level of the idea here.

Beyond compliance

Any decision someone has to sign can be built this way: insurance claims, medical triage, code review, underwriting. The question isn’t whether AI is smart enough to do the work. It’s whether you’ve given it a job where being wrong is visible.