Self-Verifying Agent Instrument Dissociates Commitment Drift from Binding Drift in Long-Horizon Agents

A recent paper on arXiv (2608.04066) presents a novel agent instrument aimed at the structural verification of long-horizon agents, tackling the issue of unreliable self-reports. This instrument incorporates a deterministic Executive that retains all beliefs, while a language model is limited to submitting typed proposals. Claims are only accepted when pre-registered predictions align with observations through code. The system disqualifies runs if per-organ write-error, render-size, or salted-canary-echo thresholds are exceeded; notably, four out of the first eight architecture runs were disqualified, each pinpointing a genuine defect. A render-invisible shadow reference compiles the plan that the complete system would have executed in every ablation cell, allowing for drift metrics even when the tested mechanism is absent. The paper concludes with a clear, singular result using this instrument, separating commitment drift from binding drift, highlighting its significance for AI safety and the verification of autonomous systems.