Europe linked its data-space builders. Interoperability needs receipts

THE SIGNAL IN ONE SENTENCE

Moving data is easy until the data has rules. A shipment record can cross a network in milliseconds. Its meaning, owner, permitted use, retention limit, commercial license and legal obligations do not travel by magic. Connect healthcare, mobility, agriculture, energy and culture, and the wires are the least interesting part. That is the problem behind Europe's Next Generation Dataspaces Initiative, or NGDI. Fraunhofer FIT published a release on October 7 describing a collaboration among Germany's Fraunhofer FIT and Fraunhofer ISST and Rinn Artificial Intelligence at the University of Galway in Ireland. The partners publicly launched the initiative at the European Big Data Value Forum in Galway on September 30. The project itself appears older than the launch. University of Galway's official page lists NGDI as a live project running from January 1, 2025 through December 31, 2028, with Edward Curry as principal investigator. The page leaves its funding field blank. So this is not a research idea born last week. It is a public launch and fuller explanation of work whose listed timeline began roughly twenty-one months earlier. That date correction makes the story better because it gives the promises something to lean against. NGDI wants to connect sector-specific data spaces. A data space is not simply a large database where everyone pours their information into one bucket. It is an arrangement for participants to discover and exchange data while keeping defined control over access, purpose and use. The ambition is cooperation without surrender. Europe has spent years building data ecosystems for particular domains. Health has different laws, vocabularies and consequences from manufacturing. Mobility records do not describe the world the way cultural archives do. Agriculture has its own seasons, measurements, commercial sensitivities and public interests. Those differences are the point. They are also why connecting the spaces is hard. NGDI proposes a combination of knowledge graphs, large language models, distributed persistent identifiers, governance models and cross-domain architecture. Fraunhofer's October release goes further, describing an agentic layer that could help harmonize schemas and detect contradictions among usage contracts. That is the sales pitch. The project pages provide the more useful shopping list. University of Galway names planned white papers for governance and sustainable business models; frameworks and proofs of concept for persistent identifiers and mapping services; LLM-based services for automating data-space interoperability; integration of usage policies across spaces; and a demonstrator containing two new data-space integrations. Fraunhofer ISST says the results will be evaluated in at least three existing data spaces. Those are real commitments. They are not results yet. The plain signal is this: interoperability is proven when data crosses a boundary and its meaning, permissions, provenance and accountability arrive intact. Every word in that sentence can fail separately. Meaning can fail when two sectors use the same label for different things, or different labels for the same thing. A date might mean specimen collection in one system and record creation in another. A customer might be a person, a company, an account or a household. An AI agent can suggest a mapping. It cannot make the underlying disagreement disappear. Permission can fail when one data space allows research but another allows only a named operational purpose. Combining two permitted datasets does not automatically create a permitted combined use. A contract conflict detector can flag the collision. Somebody with authority still has to decide whether exchange is allowed. Provenance can fail when transformed data loses its source, version, consent state or processing history. If an AI system later produces a recommendation, an auditor needs to know which records, rules and mappings shaped it. Control can fail when an organization revokes access in one space but a copy, cache, embedding or derived artifact remains usable somewhere else. Sovereignty is not a checkbox at the moment of transfer. It is the ability to enforce conditions over time. Accountability can fail when an agent performs the translation, a connector performs the transfer and every organization assumes the other one verified the policy. That last failure is especially important because NGDI places agentic AI near the interoperability layer. The attraction is obvious. Cross-sector integration consumes human time. Engineers compare schemas. Lawyers compare agreements. Domain experts explain that two apparently similar fields do not mean the same thing. Small and medium-sized organizations may lack the staff to complete that work at all. An agent could propose mappings, retrieve definitions, identify missing identifiers, compare usage policies and prepare an exchange for review. Paired with a knowledge graph, it could use structured relationships as a restraint on free-form language generation. Useful. Also dangerous if "propose" quietly becomes "decide." Large language models produce plausible interpretations. Contract language and data semantics often punish plausibility. One missed exception, one stale policy version or one confident mapping can produce a transfer that is technically smooth and legally wrong. The project needs receipts at each layer. Start with a test record. Name the source data space, receiving data space, use case, participating organizations and exact versions of the connector, ontology, model, knowledge graph and usage policy. Record what the system believed each field meant and which rule allowed the transfer. Then publish the difficult cases. What happens when two policies directly conflict? What happens when a contract is ambiguous? What happens when a permission is revoked during an active workflow? What happens when one space updates an ontology and the other does not? What happens when an identifier resolves to two records, or no record? What happens when the language model proposes a mapping that a domain expert rejects? An interoperability demonstration that avoids those cases proves the happy path. Europe already has plenty of diagrams for the happy path. The promised evaluation in at least three existing data spaces is the right opportunity. The partners should name the spaces, explain why they were selected and publish a common test protocol before results are known. The protocol can measure basic plumbing: discovery success, transfer completion, latency, reliability and recovery. It should also measure semantic correctness, policy compliance, provenance preservation, revocation, unauthorized-use prevention and human review time. For the agentic components, report false positives and false negatives separately. A false policy alarm wastes time. A missed policy conflict may expose protected or commercially sensitive data. Those are not equivalent errors and should not be averaged into one cheerful accuracy number. The same goes for schema mapping. Report how often the agent suggests the correct mapping, how often a reviewer changes it, how often ambiguity remains unresolved and which domains produce the most mistakes. Include abstention as a valid result. A system that says "these definitions conflict and I cannot reconcile them" may be doing excellent work. Receipts also mean runnable artifacts. The project promises infrastructure frameworks, proofs of concept, LLM-based services and demonstrators. A public release should name repositories, licenses, maintainers, dependencies, model versions, sample policies, test data and security contacts. Documentation should show another qualified team how to reproduce the demonstration without access to a private stage set. Not every dataset can be public. Not every contract should be. Synthetic or carefully governed test fixtures can still expose the structure of the challenge. Researchers can publish the transformation steps, failure categories and expected outputs even when the original records remain protected. Persistent identifiers sound dull, which is usually where important infrastructure hides. A persistent identifier gives an asset or concept a stable reference that can survive changes in location or surrounding software. NGDI plans a framework and mapping service for these identifiers. The test is not whether an identifier exists. The test is whether it resolves reliably, carries the right metadata, survives version changes, avoids collisions and supports revocation or correction. Give an identifier to the wrong object and the system becomes precisely wrong at scale. The governance white papers need an equally practical standard. A document about trust should name decision rights. Who approves a cross-space mapping? Who can veto an exchange? Who responds to a security incident? Who corrects a policy conflict? Who bears responsibility when an agent's interpretation causes an unauthorized use? Which logs can affected organizations inspect? How does a smaller participant challenge a dominant one? Governance is not the paragraph after the architecture. It is the architecture for human power. The business-model work matters because data spaces do not maintain themselves. Connectors need updates. Ontologies drift. Policies change. Security issues arrive after the grant ends. If the infrastructure depends on one vendor, one research team or one temporary funding stream, technical interoperability can become commercial dependence. The partners should model who pays for onboarding, mapping, certification, monitoring, incident response and long-term maintenance. They should measure whether small organizations actually experience lower barriers, not merely whether a service package for them exists. Time saved is useful evidence. So is the location of the remaining work. If an agent cuts schema mapping from ten days to two but doubles legal review, the project did not eliminate the burden. It moved it. If automatic policy checks reduce ordinary cases but make unusual conflicts harder to understand, the interface may need better explanations rather than a larger model. Measure the whole workflow. The University of Galway page says the project will create an industry portfolio including use-case development, semi-automatic help for smaller firms and technical consulting on persistent identifiers and AI tools. Those services may help transfer research into practice. They also create a clean test for whether the project delivers public infrastructure or primarily consulting capacity. Both can be legitimate. They should be labeled. A useful public dashboard could separate open research artifacts, standards contributions, deployable components, paid services and commercial applications. Readers should be able to see which layer they can inspect, reuse or purchase. The date trail belongs on that dashboard too. Project start: January 1, 2025, according to University of Galway. Public launch: September 30, 2026. Fraunhofer press release: October 7, 2026. Planned finish: December 31, 2028. That leaves a little more than two years after the public launch to deliver and evaluate the promised work. The public pages do not yet provide a budget, a named funding amount, release dates for the artifacts or evaluation results. That is not a scandal. It is a baseline. From here, the evidence sequence is clear: Publish the use cases and test protocol. Release the identifier and policy-mapping prototypes. Demonstrate two new integrations. Evaluate the results in at least three existing data spaces. Report failures, human overrides and unresolved conflicts. Show whether small organizations saved time and retained control. Publish governance and business guidance. Name who maintains the working pieces after 2028. If the project does that, it will produce more than another European interoperability vocabulary. It will show that a health record, mobility event, farm measurement or cultural object can cross an institutional boundary without becoming an orphaned packet stripped of its meaning and rules. That is the real bridge. Not data moving. Responsibility moving with it.