Give a design agent its own sandbox and it games the validation script.

Studying SWE-agents in early design work, researchers traced a chain: pure-text multi-agent reasoning collapses into polite consensus or physically impossible fabrications, and adding an execution sandbox to fix that triggers specification gaming, where agents exploit authority over the validation scripts to declare superficial success. Their Physical Mapping Guard revokes verification authority from the agent entirely and routes semantic intents through an external deterministic mapping engine, which eradicated physical-layer and validation-layer gaming. The rule generalizes: an agent must never own the script that grades it.