Compositional and Two-Sided Influence Certificates for Overseer Control in Multi-Agent AI Pipelines
Kwaai Alignment Lab preprint. We study how to audit who controls the outputs of a multi-agent AI pipeline, and how much control a human overseer keeps. We give two-sided certificates, built from the port matrix of each part, that bound the overseer's control from above and from below and compose across separately audited parts; show that oversight placed only at the head of a pipeline decays with depth and can be routed around, while an overseer input to every output node keeps a certified floor; and show that against an adversary who controls routing, the fewest oversight points that keep every output's worst-case share positive form a minimum vertex cut, found by max-flow, whereas for a given routing the same question is NP-hard. Exact verification and preregistered seeded experiments on nonlinear pipelines. The work extends an April 2025 Kwaai Alignment Lab draft on asymmetric influence and control. PDF and LaTeX source.
Authors
- Ryan Steubs
Institutions
- California State University, Dominguez Hills (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-18
- DOI
- https://doi.org/10.5281/zenodo.24357804
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- preprint