Imagine training a model to diagnose rare diseases. You blindly trust the labels in your dataset, the so-called "ground truths," as if they were objective measurements straight from reality. Wrong, according to a substantial position paper just released: every ground truth is constructed, not found. It is the visible outcome of a web of human decisions, technological tools, semantic agreements, and sampling choices that are rarely disclosed.

The central argument is that reference data are not universal but contingent. They change depending on who annotates, with which tools, in which language, and for what purpose. The machine learning community, the authors argue, would gain a great deal by ceasing to treat ground truths as indisputable facts and openly discussing their underlying assumptions. This would be the first step toward what they call "situated reliability": not a generically accurate model, but one whose limits we know, for which contexts it has been calibrated, and where it might fail.

Why self-hosted shifts the balance

The idea is not new to those working in science and technology studies, but it becomes disruptive when applied to today's landscape of Large Language Models. Companies evaluating on-premise deployments often do so for privacy or data sovereignty reasons. Yet there is an equally valid—and less discussed—reason that concerns precisely the constructed nature of reference truths.

When you fine-tune an LLM on corporate documents, you are implicitly constructing new ground truths. Using checkpoints pre-trained on generic datasets (think of those dominating public benchmarks) means importing the choices of those who defined those labels, with all their cultural and operational assumptions. A model that is "accurate" on an American dataset could perform dismally in a factory in Treviso, simply because the notion of a "machine fault" changes, local annotations reflect other maintenance practices, and measuring instruments produce readings with different tolerances.

Situated reliability calls for making these deviations explicit. Instead of chasing a single accuracy metric, the suggestion is to document the conditions of validity: who labeled the data? With what annotation guideline? Which cases remained ambiguous? Under what operating conditions does the model degrade? For a self-hosted infrastructure where the organization controls the entire stack, this transparency becomes a strategic asset, not an academic exercise.

Who loses and who gains

The discussion has economic and structural consequences. Vendors of "ready-to-use" models sell claims of universality that the paper undermines. If every ground truth is local and constructed, a commodity model promising 99% accuracy without stating which test set was used loses much of its appeal. Instead, the winners are those—often enterprises or research centers with vertical expertise—that build their own validation pipelines, curate internally annotated datasets, and invest in documenting judgment criteria.

This is not a theoretical matter: the situated reliability framework could reshape expectations in software procurement contracts. Already, some industrial tenders require not just performance specifications but also descriptions of the provenance of test data. Extending this requirement upstream to how ground truths were constructed is the next step, and the paper provides a solid argumentative basis for doing so.

The final mile of this reflection touches on sovereignty in a broad sense: the ability to define what is true in a given domain ceases to be a technical fact and becomes a political, organizational act. That is why teams choosing on-premise deployments would do well not to stop at network security or data residency. The next frontier is control over training truth: not a neutral datum, but a human construction to be governed with the same care as designing a hardware architecture.