Human-in-the-loop is the design principle that consequential AI actions require human authorisation at the moment of consequence — as an engineered gate with the evidence attached, not as an exception handler. The human is not a fallback for the system; the human is the authority the system reports to.
The quality of a HITL implementation is measured at the gate: does the approval request arrive with the full evidence file — signals, reasoning, policy, blast radius — or does it arrive as a yes/no button and a prayer? A gate without evidence is theatre.
Mature systems tier the loop: routine actions execute inside pre-authorised policy boundaries, elevated actions queue for review, and critical actions require explicit sign-off. The tiering itself is policy an operator can read and change.
This vocabulary is implemented, not aspirational — it describes how the MAIA substrate actually runs. See the platform or read the research.