Context of use is not a label. It is a contract.
The two-line 'intended use' paragraph at the front of a validation report is doing more load-bearing work than anything else in the document.
Every computational method carries a claim, stated or not. 'Predicts hepatotoxicity.' 'Ranks candidates.' 'Flags deterioration risk.' The question a regulator, a partner, or an internal review board will eventually ask is not whether the model performs — it is whether the claim was ever specified well enough to test.
A Context of Use written as a contract has parties, scope, and exclusions. The parties: who makes what decision on the method's output. The scope: which populations, test articles, assay conditions, and workflow positions the evidence actually covers. The exclusions: the indications, ranges, and situations where the method must not be used — stated explicitly, not left to inference.
The discipline this imposes is the point. When the claim is bounded, the comparator question becomes answerable: defensible against what alternative, for this decision? The reproducibility question becomes concrete: can another laboratory execute this workflow and land inside the same boundary? And the monitoring question becomes designable: which drift, in which population or upstream dependency, would invalidate the claim?
If the boundary cannot be written down, the evidence cannot exist.
Regulatory frameworks are converging on the same requirement. FDA's credibility guidance scales validation rigor to the decision's consequence, anchored on an explicit Context of Use. OECD GD 34 asks the same of alternative methods. The direction of travel is unambiguous: unbounded claims will not carry regulated decisions.
A useful habit for any team building or buying scientific AI: refuse to let a method into a decision pathway until its Context of Use reads like a contract you would sign. If the boundary cannot be written down, the evidence cannot exist.