Paper proposes self-verifying agent instrument to isolate commitment and binding drift in long-horizon agents

A new paper on arXiv proposes a self-verifying agent instrument aimed at structurally verifying long-horizon agents by dissociating commitment drift from binding drift. The instrument enforces a strict separation between proposal and execution: a deterministic Executive owns all belief and state, while a language model may only file typed proposals. Claims are admitted only when pre-registered predictions, verified by code against observations, are matched.

The instrument includes two properties that enable self-verification. First, every run invalidates itself if predefined floors for per-organ write-error, render-size, or salted-canary-echo are breached; four of the first eight architecture runs were invalidated, each localizing a real defect. Second, a render-invisible shadow reference compiles the plan the full system would have committed in every ablation cell, ensuring drift metrics remain defined even where the mechanism under test has been removed.

Using this instrument, the author reports a clean, single-variable result on a common failure in long-horizon agents: ablating the commitment mechanism flips goal-abandonment from 0.00 to 1.00 while binding error stays flat at 0.00. These results are based on three seeds per cell, up to 394 reference beats per run, and every run was gated as valid.

The paper also finds that the binding channel does not reappear as per-beat drift when its repair is ablated, because binding is code-owned and its failure class is structurally absorbed. The only residue appears upstream as a collapse in hypothesis formation.

The author pre-registers a structural defeater: task efficacy was null, with zero level completions across 52 gated runs on ARC-AGI-3. The contribution is presented as a verification methodology for agent development and the drift decomposition it makes measurable.

The paper is authored by Mohsen Arjmandi and submitted to arXiv’s cs.AI category on August 4, 2026.