A recurring pattern across the archive is the gap between stating a governance principle and possessing an operational mechanism that can change what a system is allowed to do. Risk taxonomies, ethical duties, assurance evidence and human-oversight requirements can describe expectations without identifying the authority, trigger, intervention path or residual-risk owner needed to make those expectations binding.
The review of open-source AI risk mitigation tools exposes this at the tooling layer. Evaluation and mitigation capability do not themselves determine who may accept risk or what happens when evidence crosses a threshold. *AI Agents Push Humans Out of the Loop* exposes the same issue in oversight. A human approver does not constitute a control when the system has eroded the attention, expertise or intervention capacity needed to exercise authority.
The later agent-governance reviews make the execution chain more explicit. *Designing Loyalty* translates fiduciary obligation into requirements for task-scoped delegation, conflict controls, evidence and revocation, but still requires a bridge from legal duty to execution-time admissibility. *Who Owns AI When It Breaks?* offers one such bridge through named decision owners and preauthorised stop-work powers. *The Responsibility Cascade* adds a temporal account of how knowledge and control distribute responsibility across a workflow.
The human-rights review extends this beyond a single organisation. Shared duties across developers, providers, deployers and customers remain aspirational unless they are translated into concrete constraints, monitoring, escalation and remedy. *When AI Begins to Build AI* adds an important warning: making policy machine-enforceable does not make the policy authority legitimate. *The Agentic Web Requires New Normative Infrastructure* reaches the same boundary from platform access, where proportionality remains a principle until evidence, adjudication and remedy can constrain gatekeeping decisions. The electoral-AI review reaches it again through authenticity disputes: better provenance can expose a disagreement without deciding who has authority to settle it.
Across these papers, the archive is beginning to establish a practical definition of executable governance. It requires observable evidence, an identified and legitimate decision authority, a threshold or trigger, a permitted intervention, preserved evidence of what occurred, an adjudication path when evidence or authorities conflict, and a named owner for the residual consequence. Different domains will implement those elements differently, but omitting them leaves governance advisory or moves unresolved discretion into the control plane.
The unresolved frontier is how these chains behave under disagreement. Thresholds can be contested, authorities can conflict, evidence can be incomplete, and emergency intervention can itself create harm. The archive now supports a sharper judgment: governance does not become executable merely because a control can fire; it becomes governance when the authority behind that control, the evidence supporting it, and the path for contesting it are all operationally specified. The next step for governance proposals is therefore testable operating procedures for revocation, escalation, conflicting evidence, appeal, restoration and post-incident accountability.
Traceability
This synthesis is a human-edited analytical artifact produced with AI/LLM assistance under `AI-USAGE.md`. Its claims are grounded in the constituent reviews listed in front matter and remain revisable as the archive adds implementation evidence.