Operating Across the Gap: Head-State, Visible-State, and the Discipline a Volatile Substrate Requires

Open any post-mortem from any agentic AI deployment and you will see the same framing — that head-state divergence between operator and agent is a new failure mode of large language models. This framing is technically correct and fundamentally misleading. The gap between what an operator holds in head-state and what is durably visible to a downstream partner has been named since at least Polanyi (1958) and operationally managed in semiconductor fabrication, process control, and aviation Crew Resource Management for decades. What is new is the substrate: an operating partner whose working state is volatile by architectural design, not by fatigue or training gap. The corrective discipline that prior substrates emerge naturally — durable write-down by the holder of head-state — does not emerge naturally here, because the substrate produces no internal signal that the loss is imminent. This paper documents the application of a well-mapped failure class to a new operating substrate, an empirical corpus from a small running fleet captured live, and the operator-side / agent-side corrective discipline derived from the recognition. The corpus is a single operating session (2026-06-04/05); a month of subsequent build work confirms the class recurs at both day- and month-scale, and that the corrective matures along an observable axis from operator-caught toward durable-artifact-caught.