Teaching AI to See Its Own Mistakes

type Fix =
  | { type: "move";    nodeId: string; reason: string }
  | { type: "scale";   nodeId: string; scale: number; reason: string }
  | { type: "recolor"; nodeId: string; color: string; reason: string }

The scene looked wrong and the AI didn't know.

30% of composed scenes have visible errors. The AI only has JSON — it can't see. A floating lamp. An oversized table. Parts clipping through each other.

The loop

Compose → Canvas → Screenshot → Vision Model → Structured Fixes.

The fix schema

type Fix =
  | { type: "move"; nodeId: string; position: [x: number, y: number, z: number]; reason: string }
  | { type: "scale"; nodeId: string; scale: number; reason: string }
  | { type: "recolor"; nodeId: string; color: string; reason: string }
  | { type: "remove"; nodeId: string; reason: string }
  | { type: "add"; name: string; geometry: string; position: [x, y, z]; reason: string }

Debuggable. You see exactly what the vision model thought was wrong.

$0.015

13-point quality jump (72% → 85%). 2–3 seconds. Runs automatically after compose.

The field is converging.

CMU's CADSmith (March 2026) uses a similar loop for single parts. Single parts are the first 5% — we're assembling entire machines with streaming, parallel fabrication, and collaborative editing on top.