Limits & safety net

The intelligence runs on a cloud AI accessed through a server (the game core never sees which). Only its structured actions are trusted, and only if they match the four known tools; anything else is dropped as an error. Every numeric argument is then clamped: the AI can set your tax rate only within 10–40%, and budget or sector shifts are capped at ±15 points per action, so no single command can send a value out of bounds.

Diplomacy is validated too — the target country name must resolve to a real nation, and the requested move must be one of the allowed actions, or the game refuses it. On top of that, diplomacy never concludes on the AI's word alone: proposals still require acceptance and war still requires you to pick a target. In leader-to-leader conversations there is also a per-session cap on how much a single audience can swing relations, plus a cooldown that locks relation changes for a while after they shift, so talk alone can't be farmed for endless gains.

Finally, the server layer enforces its own limits — a chat cooldown between messages and daily usage caps — so the AI is a bounded assistant, not an unlimited autopilot. The design keeps it useful and honest: it can read your true state and act on your orders, but always inside guardrails you can predict.