Definition
Whether the operator's rewrites of agent output measurably improve outcomes relative to shipping the agent's version, measured on matched samples with a randomized holdback.
The Measurement Gap
There is no metric for whether operator overrides of agent output measurably improve the final result. The counterfactual does not exist in any business record, because the unedited version was never published and therefore never measured.
Formula
OEV_c = (O_edited - O_unedited) / O_unedited, on matched samples within class c
Proposed formula, pending validation pilots.
At a Glance
Decision Supported
Whether to interrupt a session or let it run, where to set the publish threshold, and at what threshold agent-drafted communication goes out under the operator's own name.