Research

Information requirements for coordination

A central line of work studies the information that a task actually requires. It asks when a reduced representation preserves the decision-relevant content of a richer source, when collecting additional information adds cost or noise without improving the task, and how finite evidence constrains the granularity at which allocation, verification and attribution remain meaningful.

This line connects information theory, mechanism design and economic coordination. Questions originating in dispersed knowledge and economic calculation have led to formal work on sufficiency and information bounds; the resulting mathematics can then be taken back into economic problems such as incentive design and the informativeness principle.

Two papers bring the allocation, verification and hierarchical selection results together in general form. In task-relative information contracts, I ask what must cross an interface for a named downstream criterion to remain fully usable, and where a distributed protocol can lose it. The same paper shows that finer representations can need more distinctions than the available observations support. That structure also appears, independently, in the allocation and verification problem and in the hierarchical selection protocol. In task-sufficient contraction, I ask when a declared task can fix a reduced source before any encoder, rate or distortion target is chosen. The RCIT paper later named their common framework Coordination Information Theory (CIT).

In source-side sufficiency for the Information Bottleneck, I show when a summary preserves the full relevance–rate tradeoff. The interactive example removes irrelevant source bits while keeping target information unchanged. In allocation and aggregate verification, a second example shows how finer service categories improve allocation while reducing the observations available in each monitoring pool.

Adaptive, hierarchical and recursive systems

Static sufficiency is not the whole problem. In adaptive systems the relevant task, feasible actions, evidence base and representation may change over time. I study hierarchical attribution, moving information requirements, recursive learning and the conditions under which a representation should be refined rather than made more detailed by default.

Recursive Coordination Information Theory (RCIT) extends CIT and separates two requirements. An interface can keep everything a current decision needs and still omit what is needed to evaluate the situation that decision creates. The paper measures that gap and the information that recursive completion adds.

The hierarchical selection companion page follows two rounds of a small hierarchy. It shows how a selected component recovers an outcome from its allocation change, and why an inactive component must not decode the same signal.

Multi-agent systems and institutional design

Another line asks how autonomous actors should be allowed to propose actions, provide evidence, acquire authority and be evaluated after outcomes become observable. This includes reputation mechanisms, conflict handling, auditability, calibration, allocation mechanisms and work in contract theory and Austrian economics.

In TACIT, we use an rApp’s prediction accuracy to determine its influence on network decisions. Its interactive example lets you vary reputation and observe the decision change. In ORACLE, I examine how participants can independently verify the arbitration workflow; the walkthrough follows a proposal through the shared record. These papers address the decision policy and its enforcement as complementary questions.

Some projects in mechanism design and resource allocation arose independently from applied engineering problems rather than as consequences of the information-theory programme. Their relationship is therefore described explicitly rather than assumed.

Cognition and artificial intelligence

I also study attentional selection, workspace-style architectures, ensemble perception, source attribution, recursive training and synthetic data. These projects examine what information is selected, what feedback is sufficient for learning, and when provenance or confidence is a useful proxy for the task that ultimately matters.

Autonomous network management

My applied research includes O-RAN, intent-based networking, conflict detection and mitigation, calibrated automation, network attribution, privacy-preserving assurance and distributed coordination. Network-management problems have repeatedly supplied concrete settings in which more general questions about information, trust, evidence and authority become visible.

The CAMINO explanation walks through a reported scenario in which weather and traffic context affect the assessment of a network configuration proposal.

The cell-similarity explanation lets you compare performance patterns and inspect what a high correlation leaves out. The PECDAFs walkthrough shows how comparisons with similar cells can flag possible KPI degradation before a change is applied.