Every request the Architect acts on becomes a list of explicit, typed operations against your JSON Game Design Document — add a node, set a property, attach a script, bind an asset, add a mechanic, reorder nodes, set a 3D transform. Because each operation is validated before it is applied, the AI can't quietly corrupt your project: changes are structured, reviewable, and reversible.
The Architect records a snapshot of your Game Design Document before it touches anything and another once the change lands. The full history is append-only, so you can always see what a request did and roll back to any earlier state. Snapshots are tagged by what triggered them — a chat turn, a self-heal, a manual save, or a revert.
Compiling a real game project doesn't always succeed on the first try. When a build fails, the Architect can read the compile log, reason about the error, and attempt a fix — then rebuild. It's a credit-checked, bounded loop, not an infinite retry, so a failing build gets one honest repair attempt rather than silently burning your balance.
Architect actions consume credits, and the cost is tied to the work: a chat request starts around 20 credits, a compile is 10, and heavier jobs cost more. Your plan includes a monthly credit pool, and you can top up when a project hits a busy stretch. You always know roughly what a request will draw down before you send it.
Everything the Architect does lands in the same Game Design Document that powers the Scene inspector, the asset studios, and your playtest builds. Edit in chat, inspect the result in Scenes, and play it in the browser — one source of truth, no copy-paste between tools.