Narrow the claim to the evidence

A conclusion is not merely a shorter abstract. It is a claim with a target population, an outcome, an observation unit, a test procedure, and a metric behind it. Overclaim occurs when the conclusion drops those conditions and uses a word wider than the evidence. The practical response is not to impute the authors’ motive. It is to rewrite the sentence so that its scope remains visible.

Write the scope tuple before the adjective

Before accepting a strong noun or adjective, fill in a compact scope tuple: who was studied, what was labelled, what one observation represented, what data were held out, whether the model or threshold adapted, which metric was reported, and which evidence is missing. The tuple is more informative than a generic word such as “validated.” It also makes clear why two studies using the same word may be making different claims.

The rewrite habit

The following replacements are claim-narrowing templates, not universal definitions. Their exact wording should follow the study’s estimand and the evidence actually inspected.

Broad wording	A narrower wording when the evidence is limited
“a new imaging biomarker”	“a candidate image-derived feature associated with the outcome in the studied sample; biological and clinical validation were not established”
“generalisable across centres”	“performance was tested under the named internal, temporal, or external conditions; wider transport was not tested”
“clinically useful”	“discrimination was reported; calibration, decision impact, and clinical usefulness were not established”
“validated”	“tested on the specified population with the specified label, split, metric, and uncertainty procedure”
“externally validated”	“tested on an external cohort with the model and threshold [frozen or adapted] as described”

The word biomarker should not be awarded merely because software produced a feature or because a feature was associated with an endpoint. A biological, process, response, or clinical interpretation needs a validation design matched to that interpretation. Until then, “candidate feature” is not a weaker form of the same conclusion; it is a different and more accurate claim.

Likewise, generalisable needs the tested axis named. A single-site internal test does not become cross-site evidence because the model is sophisticated. A frozen external test concerns travel to the tested conditions. An external test after fine-tuning concerns transfer after adaptation. External validation is incomplete wording until the reader knows which of those procedures occurred and where the threshold was selected.

A constructed narrowing example

Imagine a single-site retrospective radiomics study that reports an AUROC but does not report calibration, decision utility, or an external test. Its abstract calls the feature a robust biomarker that generalises and is clinically useful. An honest reconstruction might say: “A candidate image-derived feature was associated with the outcome in a single-site internal test; the reported discrimination was not accompanied by external transport, calibration, or decision-utility evidence.” This sentence does not claim that the feature is useless. It says what the described design supports and keeps the unresolved questions visible.

The same discipline applies when the result is impressive. A constructed AUROC of 0.94 can remain in the sentence if its population, label, unit, and test are named. It cannot be allowed to carry the additional meanings of a validated biological marker, a calibrated risk tool, or a useful clinical intervention unless those questions were separately studied.

Missing methods are limits, not verdicts

There is a crucial difference between “not reported,” “not done,” and “done and failed.” A paper that omits its threshold-selection procedure has not proved that the threshold was selected on the test data. It has, however, left the reader unable to grant a claim that depends on a pre-specified operating point. A paper that omits calibration has not proved miscalibration; it has not shown that its scores can be interpreted as probabilities.

Reporting checklists help locate these gaps, but a checked box is not proof of correct execution, and an unchecked box is not proof of scientific failure. The reader must keep the wording proportional: “the study did not establish” is often more accurate than either “the study failed” or “the model works.”

Overclaim is often optimism, imprecise vocabulary, or a conclusion that has outgrown the original estimand. Narrowing it is an evidence-preserving act, not an accusation. The next section supplies lenses for finding omissions while keeping reporting completeness, risk of bias, and computational reproducibility as separate questions.