What the assistant can and cannot answer
It reads a summary of your timetable, not the timetable. That decides which questions it can answer well and which it is told to refuse.
The assistant does not read your .fet file. It reads a summary of it, a few hundred words long, plus the findings from the check.
That single fact decides everything below. A question the summary can answer, it answers from arithmetic somebody else already did. A question the summary cannot answer, it is instructed to refuse rather than guess.
Questions it answers well
- Who is carrying the most. Which teacher has the most hours? Teaching hours per teacher are in the summary, busiest first.
- How your file compares. Whether your class sizes, your week or your rule count are unusual against hundreds of real FET files.
- What the check found, in words. Why is this impossible? It explains the findings that are already on the page.
- Which FET rule you need. How do I stop a teacher working on Friday? It names real FET constraint types, and it checks every rule it suggests against the real list before it answers.
- Where a problem activity could go. Only when the page has already worked the answer out — see below.
Questions it is told to refuse
⛔ Where any activity can move, in general. The page works out free slots with arithmetic, for the activities that have a problem. The assistant may only quote those. Ask about anything else and it will say it does not have the answer, and point you at Find a free slot.
That refusal is deliberate and it is the most important one. A model asked to find a free slot across a whole week produces a confident wrong one, and a wrong free slot is worse than no answer: you act on it before you find out.
⛔ Anything that needs the full grid. It does not hold your timetable slot by slot, so it cannot tell you what is happening on Tuesday at 10.
Why it is built this way
A model is not arithmetic. The findings above it are proofs; the assistant is a reader of proofs.
So the design gives it as little room to invent as possible: it is handed numbers somebody else computed, it is told which questions have no answer in what it was given, and every FET rule it names is checked against the real list.
⛔ It can still be wrong. Treat an answer as a suggestion and check it against the findings, which are arithmetic and can be checked by hand.
A smaller model gets a shorter summary
A model running in your browser is given a shorter summary than a large model is. It was measured: the fuller summary made the small model worse, not better — more numbers in front of a small model are more numbers to misattribute.
So the same question can get a better answer from a bigger model, and that is expected rather than a fault.