Nine minutes is a budget, not a boast. Read every supplement paper properly and you will finish four a year. The order below exists because most papers fail early, so the cheap disqualifying checks come before the expensive ones. I use it for ward queries and for things my sister sends me at eleven at night.
The bracketed times are what each step costs once you have done it fifty times.
Funding statement and author affiliations, at the back. You are not hunting for grounds to reject. You want to know who supplied the product, who ran the statistics, and whether the sponsor could block publication.
The size of this effect has been measured. The Cochrane methodology review by Lundh and colleagues, updated in 2017 across 75 included papers, found that industry sponsored drug and device studies more often reported efficacy results favourable to the sponsor, risk ratio 1.27 (95% CI 1.17 to 1.37), and favourable overall conclusions, 1.34 (1.19 to 1.51). That is a prior to carry, not a verdict to deliver.
Red flag: an author affiliation is the ingredient supplier, and the test article was prepared there.
Find the registry identifier. Open the record. Open the history tab. You want the registration date against the date of first enrolment, and the primary outcome as originally entered.
The prevalence is known here too. COMPare, by Goldacre and colleagues in Trials in 2019, checked 67 trials published in five journals that endorse CONSORT. Of prespecified primary outcomes, 76.3% were reported correctly as primary outcomes; trials contained a mean of 5.4 undeclared new outcomes each; 58 of the 67 warranted a correction letter and 23 of those letters were published. Those are the best journals in medicine.
Red flag: registered after enrolment opened, a primary outcome edited while the trial ran, or no registration identifier at all in a 2026 paper.
Surrogate endpoint or clinical endpoint. A biomarker moved, or something happened to a person.
Red flag: the title names a disease and the primary outcome names a molecule. Also a symptom questionnaire treated as though it counted events.
Randomised n, then duration against the natural history of the thing being treated.
Red flag: n under 40 carrying a confident headline. Twelve weeks for an outcome that takes five years. Per-group numbers hidden behind a total.
Intention to treat analyses everyone as randomised. Per protocol analyses the people who complied, which means the comparison is no longer randomised, because compliance is a behaviour and behaviours predict outcomes on their own. Per protocol is a sensitivity analysis. It is not the result.
Red flag: analysed n smaller than randomised n with no flow diagram accounting for the gap. The word completers in the methods and nowhere in the abstract.
The step people skip, and the one that changes conclusions. Convert everything into events per thousand.
Take the intravenous magnesium sulphate trial in pre-eclampsia, 10,141 women, Lancet 2002. It reported a 58% lower risk of eclampsia. In absolute terms that is 0.8% against 1.9%, so 11 fewer women with eclampsia per 1,000 treated, meaning roughly 91 women treated to prevent one seizure in a condition that kills. Both numbers describe the same result. Only the second lets you decide anything.
Red flag: relative risk with no absolute counts. A standardised mean difference with no anchor to any unit a human recognises. P values in the abstract and no confidence intervals anywhere.
Red flag: before and after within a single arm, presented as an effect. Open label with a subjective primary outcome. An active comparator dosed to lose.
Find the table, not the adjective. Well tolerated is not data. For contrast, the Cochrane review of magnesium for cramps counted gastrointestinal adverse events in 11% to 37% of participants, because it bothered to count.
Read the conclusion sentence last. Then go and find the number in the results that supports it. If you cannot, you have your answer.
That totals eight and a half minutes. The remaining thirty seconds is for locating the supplementary appendix, which is where the prespecified analysis plan usually hides.
Manson and colleagues, New England Journal of Medicine 2019, the vitamin D and marine omega-3 primary prevention trial, registration NCT01169259. I use this one because it behaves well, and a clean paper teaches the checklist better than a bad one.
Step 1. Funded by the National Institutes of Health through the National Cancer Institute and the National Heart, Lung and Blood Institute among others. The vitamin D3, the omega-3 preparation, the matching placebos and the calendar packaging were donated by Pharmavite and by Pronova BioPharma with BASF. Donated product, public money, academic analysis. Noted, not disqualifying.
Step 2. Registered, with two primary outcomes prespecified: total invasive cancer and major cardiovascular events.
Step 3. Both are clinical events. Nobody's serum 25-hydroxyvitamin D concentration is the endpoint.
Step 4. 25,871 adults, men from 50 and women from 55, vitamin D3 2,000 IU daily in a two by two factorial with marine omega-3 at 1 g daily, median follow-up 5.3 years.
Step 5. Intention to treat.
Step 6. Invasive cancer occurred in 793 participants on vitamin D and 824 on placebo, hazard ratio 0.96 (95% CI 0.88 to 1.06). That is roughly 6.1% against 6.4% over 5.3 years, a difference of about a quarter of a percentage point, with an interval that comfortably includes harm. Major cardiovascular events, 396 against 409, hazard ratio 0.97 (0.85 to 1.12).
Step 9. The paper concludes that vitamin D supplementation did not lower the incidence of invasive cancer or of cardiovascular events. That sentence is exactly what the tables say, which is rarer than it should be.
Now the part that matters for how you read the news. Cancer mortality in the same trial gave a hazard ratio of 0.83 (95% CI 0.67 to 1.02). Secondary outcome, interval crossing 1. The headline that reached newspapers came from an analysis excluding the first two years of follow-up, which was not prespecified, and which the trial's own lead investigator described as hypothesis generating. The trial is not the problem here. The reading of it is.
Tables do not care what an abstract believes, which is roughly the relationship @grep/explain-analyze-has-never-lied-to-me describes between a query plan and a developer's confident opinion about the query.
One boundary on all of the above. This is how to read a paper. Deciding what to do about it is a different skill and a different conversation, usually with someone who can see your creatinine.