The Human Layer Architecture is published (DOI: 10.5281/zenodo.19120077). Paper I argued that the human layer is the critical layer in enterprise AI. This paper specifies what a real one is made of.
Five components. Decision gates: the system structurally cannot execute consequential actions without human authorization. Not "should not," cannot. Escalation protocols: defined paths for when the system meets the edge of its competence. Accountability structures: a human name attached to every consequential outcome. Override mechanisms: the human can always countermand, and the countermand always wins. Trust calibration: interfaces that keep human confidence in the system proportional to the system's actual reliability, because both blind trust and blanket distrust are failure modes.
The part that matters most, and that we suspect will be most ignored: these five are a dependency graph, not a checklist. Escalation without accountability escalates to nobody. Overrides without trust calibration either never fire or fire constantly. Gates without escalation just become queues. Implement four of five and you do not get 80% of the value; you get a system whose failure modes are harder to see.
Each component also maps to obligations that already exist in force: the EU AI Act's meaningful-oversight requirements, ISO/IEC 42001, NIST's AI RMF. The architecture is not a proposal for future regulation; it is a consolidation of present regulation into something you can actually build.
For us, March was the month the papers stopped being descriptive and started being load-bearing. The five components are now the acceptance criteria for how Timer's own autonomous capabilities ship: anything that acts consequentially inside a customer's organization goes through a gate a human owns. Slower to build. The only version worth building.