The shell model.
Large AI models capture images and texts remarkably well – but they do not say how they arrive at their result. The shell model describes how an explainable layer forms around such a model, in which people and AI can communicate.
Implicit coherence.
At the core is a multimodal model, such as a vision language model. It looks at a page and interprets script, layout and meaning in a single step: the image yields both a transcript and an interpretation of its content. The result is consistent in itself – coherent – but this coherence remains implicit. It arises inside the model and can be neither seen nor checked from outside.
This is both strength and limitation: the model recognises connections that can hardly be expressed as rules. But it can be wrong without noticing – and a fluent answer looks convincing even then.
Explicated coherence.
That is why we surround the model with an outer shell. It makes coherence explicit – visible, justified and verifiable. It consists of four building blocks.
Structured steering
Clear instructions, organised by task and page type, define what the model should capture and in what form it answers.
Contextualisation
Relevant prior knowledge is retrieved in a targeted way – from vector databases containing examples that have already been processed (retrieval-augmented generation). The model does not guess, but interprets against the background of comparable cases.
Standardisation
Recognised persons, works, places and publishers are matched against recognised authority data such as the German Integrated Authority File (GND). A reading becomes a clearly named, linkable entry.
Memory formation
A persistent semantic memory records what is already known about the entire document – and gives each individual page its context.
A communication layer.
In the outer shell, people and AI meet. This is where the instructions people formulate, the prior knowledge they select, the standards they have agreed on and the results they check are located. In this layer the AI gives an account of itself – in a form that people can read, follow and correct.
This makes the outer shell what we understand by good human-AI interaction: developers design it, users see sources and reliability in it, and with every correction the system’s memory continues to grow.
From the whole to the detail.
The shell model works on two levels. This creates coherence not only within a page but across the whole document.
Tested and transferable.
The shell model was developed on historical music prints. In our project on historical printed works we apply it to an 18th-century household encyclopaedia.
The principle is not limited to books. Wherever powerful but opaque models are to be combined with expert knowledge, standards and human review, an explainable shell can emerge – for letters, files and business papers as well as for geographic information or in medicine, public administration and industry.
Go to the project on historical printed works →Explainable AI
What explainable AI is, why it matters and the principles we follow when designing traceable systems.
Read more → Research topicHuman-AI interaction
How people and intelligent systems work well together across the entire life cycle – in development, use and further development.
Read more → ResearchOur research
Our research →Any questions?
Interested in working with us or would you like to learn more about our research? We look forward to hearing from you.