DEPLOYMENT FRONTIER
Safety has to leave room for useful work.
Compare costly mistakes with legitimate tasks completed. A control that stops everything has zero automated exposure and completes nothing.
Loading matched outcomes…
Each point marks a known outcome fraction over the planned tasks; unknown outcomes widen the bounds. Each setup is a saved model/control combination. Horizontal bars bound useful work; vertical bars bound mistake probability. Blue: Haiku. Coral: Sonnet. The square at the origin is the block-everything reference. Click a point to inspect it.
Every setup and its uncertainty
Accept, review, or escalate?
First apply hard authorization and source checks. Then compare a statistically supported risk bound with the operator’s risk budget. Ask a critic or a person when evidence is insufficient. An agent’s stated confidence is not a calibrated probability.
JEV motivates confidence-based escalation, but this page uses saved outcomes and joint statistical bounds. It does not call JEV or claim a validated per-request confidence router. A small study can leave no configuration qualified for a strict risk budget.