What it computes

A graph neural network computes each element’s state from its own attributes and the states of its neighbours, and repeats the step, so that information moves across the graph. After a few rounds a node’s state reflects elements several steps away. On an enterprise’s graph this lets the effect of a change in one place be estimated several dependencies further on, where nobody documented the chain: a provider’s capacity decision surfacing as a wrong claims decision three services downstream; a duplicate in a counterparty master surfacing as a double entitlement.

Three jobs

Over the Petri structure the network is specified to do three things. Estimate the parameters no record states, learning from the enterprises already modelled how an element of a given kind behaves. Forecast the consequence of a disruption where no one documented the dependency. Attach to every estimate a range with a stated coverage guarantee, so that a figure resting on thin evidence carries a visibly wide range. Every such estimate enters a report as an inference, never as a citation.

Many curves, not one

An enterprise does not have a single resilience curve. Each commitment can be reached by many paths through the estate, and each path fails differently under each scenario, so an obligation carries hundreds or thousands of curves. In the present specification a simulator produces them; the network’s role is to estimate them faster, and for parts of the graph whose parameters are unknown.

The learning loop

Every model built improves the shared assets: what is learned about how enterprises of each kind are built is shared across every model, and no client’s data crosses to another. The network learns from models whose true structure is known, so it begins once the first corrected engagements exist and improves with each one.

Status

The learned layer is the direction of the build, not yet in the specification. The first two states run today on established methods: statistical pooling for missing values, a scenario simulator, and a reliability test with a pricing step. The pairing of Petri-net structure with graph neural networks entered the published literature in 2025 as a general architecture; the firm’s search found no commercial application to an enterprise’s resilience.