Long-running behaviour
How an effective operating policy can move over time without one obvious failure event.
Research / VaHive Systems Lab
What we research, and why.
We study governance architecture for agentic systems: how authority, model behaviour, execution, evidence, and time interact when a system acts beyond a single session.
The premise
The model is not the whole system.
Agentic systems work across prompts, memory, permissions, tools, policies, external services, and human review. A research programme that only observes one of those layers cannot explain every failure mode created by the full system.
Authority & boundaries
How specified limits, permissions, and pre-execution records can structure consequential actions.
Representation gaps
How input, stated reasoning, activations, and execution intent may be different windows onto one event.
Evidence in practice
How an operator can inspect changes and decisions rather than rely on a retrospective story.
Published preprints
Research record
These are preprints. They propose frameworks and architectures; they are not certification, independent validation, or proof that a mechanism is deployed.
14 March 2026
MAGUS v3.0: A Governance Architecture for Structural Alignment Drift in Long-Running Agentic AI Systems
A theoretical governance architecture and open issues register for structural alignment drift in long-running agentic systems. It is not a report of a deployed product.
19 May 2026
Connecting Activation Geometry to Execution Intent: A Multi-Representation Framework for Detecting Computational Divergence in Agentic AI
A proposed framework for comparing input, expressed reasoning, activation-level representations, and execution intent. It makes no empirical performance claims.