Explanation Structure Model (ESM) | Umnai

An Explanation Structure Model, or ESM, organises the information needed to produce explanations from an Umnai model.

An ESM starts from the predictions and raw explanation data produced by an Explainable Neural Net (XNN), then structures that information so it can be inspected, transformed, filtered, ranked, summarised, or used in further reasoning.

ESMs make the explanation process explicit. They describe how answers, justifications, symbolic links, dataset concepts, rules, bins, decompositions, and activation paths fit together to produce an explanation that can be presented to a user or consumed by another system.

Why ESMs matter

An XNN produces a prediction and the raw explanation data behind it. An ESM provides the structure that turns that raw material into a practical explanation.

Given a set of answers, such as predictions, and a set of justifications, such as the workings and decisions that produced those answers, an ESM specifies what should happen next.

For example, an explanation may need to be:

Step	Purpose
Summarised	Reduce detailed explanation data into a clearer form.
Aggregated	Combine related explanation signals.
Filtered	Select the most relevant information.
Ranked	Prioritise the most important explanation components.
Reasoned over	Apply additional logic to reach a final explanation or decision.

This makes the chain of reasoning explicit. Another way to think about an ESM is as a plan of action for producing an explanation.

ESMs as hypergraph models

ESMs are hypergraph models that store information about the symbolic links inside an XNN, the symbolic concepts associated with the underlying dataset, and other symbolic knowledge used by the explanation system.

They package XNN outputs into an accessible structure that can include:

Information	Description
Rules	Symbolic rules associated with model behaviour.
Bins	Feature segments used to define rule and partition boundaries.
Decomposition values	Feature-level components of interaction attributions.
Activation paths	The activated structures that contributed to a prediction.
Dataset concepts	Symbolic concepts and metadata associated with the underlying dataset.

When an ESM embeds an XNN, it creates the neuro-symbolic transforms and graph connections needed to enrich the raw explanation data from the XNN. This is what enables Hybrid Intelligence explanations to be produced from the model output.

Components

An ESM contains several layers of information and processing.

Component	Role
Explainable machine learning pipeline	Performs pre-processing and post-processing to transform input data into formats suitable for neuro-symbolic processing, and back again where needed.
Embedded neuro-symbolic machine learning model	The XNN that produces predictions and raw explanation data.
Neuro-symbolic hypergraph	Stores symbolic links within the XNN, including symbols, rules, and bins.
Dataset metadata	Stores dataset information needed for explanations, including properties, statistics, histograms, anomalies, and mappings.

Together, these components allow the ESM to connect model behaviour, dataset meaning, and explanation structure.

From model output to explanation

ESMs start with information from their embedded XNNs.

An XNN provides specific information about the problem it is solving: predictions, module attributions, feature attributions, activated partitions, rules, bins, and decompositions. The ESM uses that information and applies additional layers of knowledge to make the explanation broader, more structured, and more useful.

These layers can include:

Layer	Example
Cause-and-effect information	Add causal or domain relationships to the explanation.
Filtering	Select the most relevant explanation components.
Ranking	Prioritise the most important information.
Summarisation	Present information in a form that is useful to the people consuming it.
Further reasoning	Support a final decision, such as a go/no-go decision.

This allows an ESM to move from detailed model behaviour to a higher-level explanation, and from broad explanation structures back down to specific situations when needed.

Multiple models and chaining

ESMs can combine multiple XNNs.

They provide a framework for XNN chaining, multiple-model systems, mixture-of-experts approaches, and multi-agentic systems. This allows explanation structures to span more than one model when the system needs to coordinate several sources of predictive or symbolic information.