The LLM Proposes, the Executive Disposes: A Self-Verifying Agent Instrument that Dissociates Commitment Drift from Binding Drift in Long-Horizon Agents
06:00 · August 6, 2026 · arXiv cs.AI RSS
How do you verify a long-horizon agent when its own state and self-reports are exactly what you cannot trust? We present an agent instrument built so that verification is structural rather than post-hoc. A deterministic Executive owns all belief; a language model may only file typed proposals, and a claim is admitted only when a prediction pre-registered before acting is matched against observation by code. Two properties make the instrument a verifier of its own science, not just of the agent: every run invalidates itself when per-organ write-error, render-size, or salted-canary-echo floors are breached (four of the first eight architecture runs were invalidated, each localizing a real defect); and a render-invisible shadow reference compiles the plan the full system would have committed in every ablation cell, so drift metrics are defined even where the mechanism under test has been removed. Using this instrument we report a clean, single-variable result on a failure every long-horizon agent suffers: ablating the commitment mechanism flips goal-abandonment from 0.00 to 1.00 while binding error stays flat at 0.00 (three seeds per cell, up to 394 reference beats per run, every run gated valid). The binding channel, by contrast, does not reappear as per-beat drift when its repair is ablated -- because binding is code-owned, the failure class is structurally absorbed, its only residue appearing one layer upstream as a collapse in hypothesis formation. We report these under full disclosure that task efficacy is null (zero level completions across 52 gated runs on ARC-AGI-3), pre-registered as a structural defeater. The contribution is a verification methodology for agent development and the drift decomposition it makes measurable.
Summary
The paper presents a self-verifying agent architecture designed to isolate two distinct forms of goal drift in long-horizon LLM agents. A deterministic Executive maintains all belief state and admits only typed proposals from separate model organs—an Observer, Surveyor, and Actuator—each of which must pre-register predictions before any action occurs. A claim enters the agent’s state only when deterministic code matches the logged prediction against subsequent observations, preventing the model from authoring its own verification.
Two structural safeguards make the instrument self-auditing. Every run is automatically invalidated if per-organ write-error rates, render sizes, or salted-canary echo floors are breached; four of the first eight architecture runs were rejected by these gates, each exposing a concrete implementation defect. In parallel, a render-invisible shadow reference compiles the plan the complete system would have produced in every ablation cell, supplying defined drift metrics even when the mechanism under test has been removed.
Experiments conducted on ARC-AGI-3 with three seeds per cell and up to 394 reference beats per run demonstrate a clean dissociation. Removing the commitment mechanism raises goal-abandonment from 0.00 to 1.00 while binding error remains at 0.00. Ablating the binding repair produces no per-beat drift increase; because binding is enforced in code rather than attention, the failure manifests upstream as reduced hypothesis formation. All 52 gated runs completed zero tasks, a null result registered in advance as a structural defeater.
The work therefore contributes a verification methodology rather than task performance. By externalizing commitment and binding as independent, code-owned switches and by enforcing run validity and shadow references at the architectural level, the instrument renders drift measurements reproducible and attributable even when the agent’s own state cannot be trusted. Full logs and pinned code are released for independent replay.
Why it matters
High technical depth, reproducibility, and novelty in agent verification directly support Dutch/EU priorities on transparent, ethical AI; Dutch researchers and advanced teams can apply the instrument and drift taxonomy to reliable agent development.
