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.