Semantics Is the Part Everyone Keeps Skipping

by Marina VasquezHead of Semantic Systems, HMND

The last week’s movement around federated data access keeps running into the same wall: teams want cross cohort decisions before their data can agree on what a subject, a sample, an assay, a protocol, or a product even is. That is why federated access keeps stalling across preclinical, translational, clinical, and manufacturing work, even as standards programs, controlled vocabularies, and private linkage methods keep getting sharper.

What people are really frustrated by

The frustration is not that data is locked away. It is that data is reachable but semantically unreliable. CDISC gives clinical and non clinical research a standards spine, including foundational models, exchange standards, and therapeutic area standards, but those structures do not magically reconcile a preclinical entity model with a clinical one or a manufacturing record with a translational biomarker table. Federated learning and analytics are being pitched as a way to use more diverse datasets without aggregating them, but the same literature still points to privacy constraints, distinct silos, and the need to operate across distributed stores rather than pretending they are one coherent system.

That is the real irritation. People are being asked for a unified answer while the underlying systems are still speaking different dialects. The fight is not really about access. It is about meaning.

Why adoption is hard across functions

Adoption breaks when each function has different data gravity. Clinical groups are pulled toward submission standards, traceability, and controlled terminology because those are the rules of the road for regulatory exchange. Manufacturing teams are dealing with process control, product traceability, audit trails, and the need to connect process inputs, outputs, and outcomes across sites and contributors. Translational teams are trying to bridge biology, assays, and evidence packages without flattening nuance into a brittle schema. Preclinical teams often sit on models that are locally useful but globally misaligned with downstream identifiers and study structures, which means even a simple cross cohort question becomes a reconciliation problem.

The hard part is governance, not software theater. Standards only work when ontology governance is real, controlled vocabulary is maintained, and entity resolution is treated as an operating discipline rather than a one time mapping exercise. Transcelerate, CDISC, and FDA sit inside the larger standards gravity around clinical data exchange and submission readiness, but that alignment does not arrive by announcement. It has to be enforced through shared definitions and maintained over time.

The WEF’s federated consortium framing makes the same point in plainer language. A federated system only fits when there is a clear problem that distributed access can solve, and when the model can navigate different data types, ownership rules, and local controls. That is why teams stall in the real world. The architecture is rarely the hardest part. The hard part is deciding which meanings are stable enough to federate in the first place.

Where queryability gets mistaken for understanding

Failure happens when a platform can answer a query but cannot justify the answer’s meaning. That is the quiet trap. You can federate access, run a cross site search, and pull records from multiple nodes, but if the identifiers are inconsistent, the ontology is weak, or the provenance chain is thin, the result may be machine readable without being decision ready.

This is where teams confuse queryability with understanding. Queryability says the system returned something. Understanding says the returned thing is the same entity, in the same context, under the same definition, with the same interpretive rules. Without that, cross cohort decisions become a kind of organized self deception. The dashboard looks alive, the comparison looks clean, and the operational consequences are still built on mismatched semantics.

That failure shows up as duplicated subjects, broken lineage, incompatible endpoints, and analyses that look federated but are actually stitched together by hope. It also shows up as governance fatigue. Once the second or third function discovers that its terms do not survive translation intact, trust collapses.

The real leverage

Semantics is what lets federated access become more than distributed retrieval. It is what makes standards actionable, ontology governance durable, and entity resolution credible across domains that were never designed to agree. The teams that keep asking for cross cohort decisions are not wrong. They are early. They are asking for a decision layer that the data layer has not yet earned.

And that is the point. Federation without semantic discipline is just a wider corridor for ambiguity. With it, the same corridor becomes navigable.

If you are living inside this problem, it is usually more useful to compare notes than to pretend the stack is cleaner than it is. Semantics is leverage, not decoration, and the people who have to make it work are rarely the ones who got to choose the original model.