%% You should probably cite draft-sato-soos-hem-07 instead of this revision. @techreport{sato-soos-hem-05, number = {draft-sato-soos-hem-05}, type = {Internet-Draft}, institution = {Internet Engineering Task Force}, publisher = {Internet Engineering Task Force}, note = {Work in Progress}, url = {https://datatracker.ietf.org/doc/draft-sato-soos-hem/05/}, author = {Tom Sato}, title = {{The Human Escalation Mechanism (HEM) for Agentic AI Systems}}, pagetotal = 47, year = , month = , day = , abstract = {An AI agent that has been authorized to act autonomously has no inherent mechanism to stop itself. If its mission requires a decision that exceeds its authorization, if policy mandates human judgment before proceeding, or if the agent itself reaches the boundary of its reliable competence, what happens? Without a protocol specifying the answer, one of three failure modes occurs: the agent proceeds beyond its authorization and executes actions that no human approved; it stalls silently with no notification to any principal; or it continues running under a mission that has already entered a terminal state, producing actions with no legitimate purpose. In all three cases, the humans responsible for the system find out too late. This document defines the Human Escalation Mechanism (HEM): a normative protocol specifying what a Governance Execution Controller (GEC) does when an AI agent session requires human judgment before execution may continue. HEM replaces the three failure modes above with a single governed path: the GEC places the session into a formally defined HEM\_PENDING state, routes a structured escalation request to one or more designated human principals along an ordered designation chain, enforces a prohibition on all state transitions until a human decision is received, and processes six defined human decision types. HEM also defines the Policy Rationale Declaration (PRD), which links Cedar policies that route to HEM with machine-readable rationale, and the Decision Rationale Record (DRR), which captures the human principal's reasoning for audit and learning purposes. Version -05 adds ten new HEM interaction classes (HEM-PRE-1, HEM-PRE-2, HEM-DS-1, HEM-DS-2, HEM-LIM-1, HEM-DIV-1, HEM-HIGH-1, HEM-FAT-1, HEM-EMO-1, and HEM-CONSENT) with full normative specifications, trigger conditions, GAR ALE registrations (ALE-030 through ALE-041), and five new Security Considerations addressing the HEM channel attack surface. INV-HEM-01 (The Surfacing Obligation) is added as a KernelSpec invariant. Section 7 (Human Readiness Score) and Section 8 (Tier 0-A Integration) are new normative sections. HEM is enforced by the GEC, not by the agent and not by the application layer. An agent cannot opt out; an application cannot suppress it. This non-bypassability is the source of HEM's regulatory utility and provides the technical specification for human oversight required by EU AI Act Article 14.}, }