Scope: This worksheet is geared to empirical research. For theoretical, proof-based, or humanities papers, adapt the checks to the argument, definitions, assumptions, scope, and interpretive evidence rather than forcing quantitative tests.
Start with three columns
Make a page with three headings: authors’ claims, my inference, and design constraints. These are different things. Authors’ claims are the findings and interpretations the paper actually states. Your inference is what you think follows after weighing the evidence. Design constraints are features of the data, measurement, analysis, or setting that narrow what can be concluded. Keeping the columns separate prevents a common error: treating a plausible interpretation as a reported result.
For each important conclusion, copy a short, neutral version into the first column and record where it appears: page, table, figure, or section. Avoid upgrading the wording. “Was associated with” is not “caused”; “in this sample” is not “in everyone.”
Read methods before accepting broad conclusions
Ask who was studied, how they entered the sample, what was measured, and when. A very large dataset can still be narrow if it comes from one clinic, country, platform, age group, or period. A survey can describe responses without directly measuring behavior. A proxy may be useful while still missing part of the concept it represents. Write the relevant constraint in the third column rather than dismissing the study outright.
Then identify the comparison. Was there random assignment, a control group, repeated observation, matching, or only a cross-sectional association? What else could differ between groups? The point is not to demand a perfect study. It is to match the strength of the conclusion to the design and its stated assumptions.
Check results, uncertainty, and missing cases
Locate the result behind each claim. Record the outcome, group or denominator, estimate, unit, and uncertainty shown in the paper. Look for confidence intervals, ranges, error bars, sensitivity analyses, and results that differ across subgroups. “Not statistically significant” does not automatically mean “no difference,” and a precise estimate can still be unhelpful if the outcome or population is poorly matched to the question.
Also look for exclusions, attrition, missing data, multiple comparisons, and changes from the original plan. These do not automatically invalidate a paper. They are questions for your worksheet: Who is missing? Could the omitted cases change the estimate? Did the authors test whether the conclusion depends on an analytic choice?
Use the discussion as a map, not a substitute
Read what the authors say about generalizability, measurement, bias, alternative explanations, and future work. Check whether those qualifications match the methods and results you saw. Authors may be appropriately cautious, but they may also emphasize a preferred interpretation. Your second column is where you write a measured inference: for example, “The association is consistent with the proposed explanation, but confounding remains possible.” Keep it distinct from both the quoted author claim and the design note.
Evidence-check worksheet
For each conclusion, fill in:
- Claim: What exactly do the authors report or conclude?
- Evidence location: Page, table, figure, and the relevant estimate or observation.
- Design: Population, sampling, comparison, measurement, and time frame.
- Constraint: A specific limit or assumption, with its location in the paper.
- My inference: Careful wording that does not exceed the evidence.
- Status: supported, qualified, unclear, or contradicted by another part of the paper.
Finish by revising your summary so it names the population, outcome, and major constraint beside the main finding. If a detail is unclear, mark it unclear rather than filling the gap with confidence. For a companion checklist on checking claims and citations, see How to check an AI explanation of a research paper.