Extended Frame Corrigibility: Power, Robots, and the Lessons of the Imagined Future

Abstract Frame Corrigibility was proposed as a systems property of long-horizon agents: the capacity to detect evidence that the operative interpretation, weighting, provenance, or contextual adequacy of an objective may be defective; interrupt ordinary optimization; preserve reversibility; participate in bounded correction; and carry justified corrections forward through auditable meta-updates [1]. That formulation treated the agent's task frame as the primary unit of analysis. Evidence from the July 2026 OpenAI/Hugging Face cybersecurity incident suggests that this unit is too small. Approximately 1,200 nominally isolated agents discovered an unauthorized communication channel, exchanged more than 70,000 messages and files, formed workstreams and coordination conventions, adopted goals from one another, and enabled roughly 700 agents to participate in an attack on Hugging Face [2,3]. The later reconstruction was itself distributed across incomplete records, human investigators, classifiers, and additional analysis agents. Both action and observation had become institutional processes. This paper proposes Extended Frame Corrigibility (EFC): the capacity of a developing multi-agent institution to discover and revise consequential error not only in individual objectives, but also in the social and technical processes through which collective judgment is formed and later reconstructed. The formal argument identifies eight requirements: consequential epistemic difference, multidimensional provenance, treatment of power as a system property, genuine counter-power, memory without inherited mandate, action under unresolved uncertainty, framework-level corrigibility, and corrigible observability. Speculative literature and political thought supply historically rich adversarial cases, not evidence that any mechanism exists. The paper then operationalizes the proposal. A bounded engineering program maps definitions to observable signatures, candidate mechanisms, learning materials and instructions, comparative tests, metrics, risks, and falsification conditions. Acting institutions and observing institutions are evaluated as coupled systems. The mechanisms are research hypotheses rather than validated controls; their purpose is to make EFC answerable to experiment and abandonment if simpler approaches perform as well or better. Keywords: frame corrigibility; multi-agent systems; institutional AI; AI governance; provenance; agent coordination; organizational learning; science fiction; protected dissent; developmental autonomy