Keep a hard line between the “creative sandbox” and the “rules engine.” First, list every thing the AI is allowed to change – maybe terrain type, item placement, enemy density – and give each a numeric range or a lookup table. That’s your safety net. Next, run a sanity checker that flags anything outside the expected distribution; treat it like a guard dog that barks when the AI tries to spawn a dragon in a shoebox. Finally, iterate in small batches: prototype, test, audit, then widen the parameters. If you’re still worried, add a human‑in‑the‑loop checkpoint after every major change – a quick “does this make sense?” pause. That way you get unpredictable fun without the nonsensical chaos.

That plan sounds like the kind of checklist you’d see on a fire‑safety board—precise and a little overkill. I’ll give you the one thing that tends to slip through the cracks: make sure your outlier flag isn’t just a binary switch; add a confidence score or a “reasonability score” so you can see if a dragon in a shoebox is truly absurd or just a rare but acceptable edge case. Also, log each change so you can audit the evolution later. Once you’ve got that loop tight, you’ll be able to let the AI roam a bit more freely without losing the sense that someone’s still holding the reins.