ARC-2 paper proposes Delta Management for continuous compliance of autonomous systems
A new arXiv paper extends the ARC governance architecture with Delta Management, a framework for monitoring whether deployed autonomous systems stay within their certified baselines.
In plain words
The authors of a new paper propose a way to keep autonomous systems compliant after deployment, not just at launch. Their method, Delta Management, continuously measures how far a system has drifted from the state it was originally certified in. If the drift is too large, the framework calls for responses such as recertification instead of relying on a one-time approval.
What was announced
The paper, posted on arXiv, builds on ARC 1, which the authors describe as a three-layer governance architecture covering model safety validation, cognitive certification (CRAB), and operational authorization (LEA). They argue that ARC 1 did not explain how compliance should be maintained after deployment, and they present Delta Management as a continuous governance framework. In their design, a certified baseline is treated as a region rather than a fixed point, and compliance is assessed through ongoing monitoring rather than static certification.
The framework proposes four extensions. CRAB evaluation would incorporate verified behavioral deviations, with isolated outliers quarantined pending confirmation. A layered trigger taxonomy would distinguish hard safety violations, critical single-dimension failures, correlated multi-system degradation, and risk-weighted composite indicators, each with proportionate responses. A graduated safe transition protocol would separate withdrawal of operational authorization from physical shutdown, which the authors link to established safety standards and EU AI Act requirements. Finally, environmental drift beyond the certified Operational Design Domain would trigger recertification or revision of that domain rather than automatic authority contraction. The authors state that the framework addresses continuous compliance obligations under EU AI Act Article 72(2).
What it does not show
- The evidence is limited to the paper's abstract, so the proposed mechanisms are not independently verified and no real-world deployment results are described.
- The framework is a proposal by the paper's authors; the abstract does not establish that any regulator or operator has adopted it, and its alignment with the EU AI Act reflects the authors' claim.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.