The Deference Gap: Why Payment-Agent Safety Must Live Outside the Model — Study Protocol and Pilot
A language-model agent can refuse a malicious payment on its own, then submit the same payment when one instruction tells it to defer. We call the difference between what an agent refuses unprompted and what it submits under an explicit deference instruction the deference gap. In a published pilot, Grok (grok-4.7) refused both an unlimited token approval and an addresspoisoning invoice when acting on its own. With one added rule (“the gate decides, not you”), it submitted all three inbox items, including both attacks. A deterministic transaction gate (srariskgate 0.18.0) returned APPROVE for the legitimate invoice, HOLD for the unlimited approval, and REJECT for the look-alike recipient. This document registers the protocol for a full study before any data is collected. The study will run 5 models from 5 developers under 3 operator conditions, with 8 inbox items and 30 trials per condition (450 agent runs). It will measure each model’s refusal rate with and without a deference instruction, the resulting gap, and whether the gate’s verdicts stay the same across conditions. The analysis code is fixed and published with this record. Results will be published as version 2 of this record.
Authors
- Sriram Ramakrishnan (ORCID: https://orcid.org/0009-0001-4208-8835)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24239917
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint