Executor AI and Witness AI: An Independent Cognitive Oversight Architecture for Highly Autonomous AI Systems

This paper proposes Executor AI–Witness AI (EWA), a conceptual architecture for supervising highly autonomous AI systems. It separates the capacity to act (the "Executor") from the capacity to supervise (the "Witness"), which observes, evaluates risk, and can halt or escalate to human review — without operational autonomy of its own, and without knowledge of the Executor's internals or vice versa. The architecture was hardened through eight adversarial review scenarios and is compared against existing work (AI Control, untrusted monitoring, distributed-threat monitoring). It includes seven falsifiable hypotheses and reports results from small experimental pilots — including preliminary and null findings — reported transparently, together with a call for community follow-up.Version 2 (September 2026). Revised version. Corrects the stated direction of hypothesis H1; extends the adversarial review from eight to nine scenarios, adding Witness availability as an attack surface; adds specifications for the Witness runtime invariants, the pre-execution gating boundary, a reversibility-tiered human review quorum, and a candidate multi-agent attribution scheme; adds an appendix specifying the safety principles as an artefact; declares the paper's scope explicitly as a position paper accompanied by a reproducible harness; includes complete, runnable supplementary code for the first time, together with the task data and a cross-lab protocol pairing a Gemini Witness with a Claude Executor; and discusses the agent coordination incident reported by Reuters on 4 September 2026.