Models need identifiers, not just papers
Composite computational models are exposing the object we still don’t identify well enough, writes Joshua Brown
We’ve spent decades getting better at scholarly identity. Papers have DOIs. Researchers have ORCID iDs. Institutions have ROR IDs. Datasets can be deposited and cited. Source code can be archived with identifiers such as SWHIDs that resolve to a specific software artefact. Then a computational model crosses from one project into another, becomes an input to something larger, and the downstream researcher can still be left asking a very basic question: which model, exactly, did I just inherit?
That question is becoming harder to avoid because the model is no longer always a thing sitting politely behind a paper. In complex computational research, one model’s output can become another model’s input. A segmentation model can feed a mechanistic model. A mechanistic model can feed a risk model. A digital twin can combine imaging, biomarkers, longitudinal observations, software components and assumptions that live on different time scales. The paper may explain the work brilliantly and still fail to identify the exact scientific object travelling through the chain.
When model identity becomes an interoperability problem
Virtual human twins make the problem unusually visible. The European Commission’s 2026 Cancer Mission call asks researchers to build dynamic, multiscale VHTs using large volumes of FAIR multimodal data, update and validate them with longitudinal patient data, and make data and model assets available through UNCAN.eu and the Advanced VHT Platform. That’s exactly the kind of research architecture in which weak model identity stops being a documentation nuisance and becomes an interoperability problem. If models are meant to move, combine and be reused, their scientific lineage has to move with them.
A citation can tell me which paper described a model. It can’t necessarily tell me whether I’m looking at the model before or after retraining, which parameter set generated a particular result, what preprocessing pipeline sat upstream of it, which validation evidence belongs to that version, or whether a later update crossed a boundary that should have triggered revalidation. The distinction sounds technical until the model becomes somebody else’s dependency. Then it’s provenance.
Giving the model a passport
I’d give the model a passport. I don’t mean another badge, another PDF, or another compliance document somebody fills out once and forgets. I mean a machine-readable record that travels with the model output and binds identity to scientific state. At minimum, it should carry a persistent model and version identity; the provenance of inputs and transformations; the state variables and time scale the model represents; assumptions and validity boundaries; uncertainty and sensitivity information; dependencies on upstream data, software and models; validation evidence, including failed or negative tests; licence and reuse conditions; and the conditions under which an update requires revalidation.
The phrase isn’t mine alone. A 2025 health-AI paper proposed an AI Model Passport as a structured, machine-readable traceability framework and implemented it with provenance standards and MLOps tools. That work matters because it shows the concept can be operational rather than rhetorical. But the research-information problem is wider than AI. Mechanistic models, simulation models, hybrid models and composite digital twins need identity too. The useful move now is to connect model-level identity to the scholarly infrastructure we already have instead of pretending we need to invent an entirely separate universe.
That means the passport doesn’t have to replace DOIs, repository records, software identifiers, Git commits or dataset identifiers. It can bind them. A DOI can identify a deposited release. A SWHID can identify a software artefact. A dataset can carry its own persistent identifier. The passport’s job is to say which of those objects were assembled into this model state, under which assumptions, with which evidence, and what changed when the model moved from version A to version B. The identifier tells me what object I have. The passport tells me what scientific state that object represents.
Publishing the model as a research object
This is where publishing practice has to catch up. We still often publish the paper as the primary object and treat the model as supplementary material, code, or a repository link. That order made sense when the model mainly existed to support the paper. It makes less sense when models are intended to be reused directly by other researchers, plugged into shared infrastructure, or combined into larger systems. The model is becoming a research object in its own right. It needs citation, versioning and provenance at the same level of seriousness we now expect for datasets.
There’s also a simple rule publishers and repositories could adopt: a result should resolve to the model version that produced it, not merely to the latest version somebody has uploaded. ‘Latest’ is convenient for software distribution. It can be disastrous for reproducibility. If version 1.3 generated the published result and version 1.5 changed the calibration data, uncertainty treatment or validity boundary, a reader shouldn’t have to reverse-engineer that history from a release log.
Recording where confidence ends
The same applies to failure. We’re very good at attaching positive validation claims to models. We’re less consistent about preserving the tests a model failed, the populations or conditions in which it stopped behaving well, and the interface assumptions that broke when it was connected to something else. Those failures are part of the model’s identity because they define where the model should stop travelling. A passport that records only what worked is advertising. A passport that records where confidence ends is research infrastructure.
FAIR principles are essential, but findable and reusable aren’t synonyms for intelligible inheritance. A model can be technically available and still be practically opaque if the next researcher can’t reconstruct what version produced an output, which assumptions travelled with it, or whether uncertainty disappeared at an interface. Open access gets the object through the door. Provenance tells us what came through with it.
Model identity is scholarly communication
That’s why I think research information systems should stop treating model identity as somebody else’s software-engineering problem. The moment models become citable, reusable, federated research assets, identity becomes part of scholarly communication. Cross-project infrastructure shouldn’t merely ask, ‘Can these models connect?’ It should ask, ‘Can the receiving researcher tell what he has inherited, what changed, what failed, and when the evidence no longer travels with the model?’
The next generation of computational research is going to be full of systems made from other systems. We can either make lineage explicit now or spend the next decade reconstructing it after the fact. We already learned that papers, researchers, institutions, datasets and software need durable identity. Models are telling us they do too.
Joshua W.J. Brown is a Toronto-based independent systems researcher and writer working across computational modelling, provenance, interoperability, validation and research infrastructure