Trial tables bring together participant counts, percentages, treatment results and measures of uncertainty. Those numbers are later reused in evidence reviews and clinical guidelines. When values do not agree, they deserve a closer look—but the first task is to distinguish a problem in reading the table from a contradiction in the report itself.
I examined numerical consistency in published randomized-trial reports and worked on connecting those checks to evidence appraisal. A program’s flag should not automatically become an error judgment was the principle around which I rebuilt the manuscript.
Follow a flag back to its source
Automated checks can start with relationships governed by explicit rules, such as the link between counts and percentages. Yet the denominator might be stated elsewhere, values might be rounded, or extraction might have displaced a table’s rows and columns. Listing the calculations without returning to the original table and notes can mix unresolved signals with genuine contradictions.
I connected the checked items to source locations and organized the working record to show what supported each judgment. Even after identifying a problem, further questions remain: is one corrected value uniquely supported, and would that correction change the interpretation of the study?
A possible correction may still need a question
Finding a contradiction does not automatically reveal the correct value. If several repairs are possible, choosing one would add an unsupported assumption. I kept those cases open for clarification and distinguished reporting problems from problems that could affect interpretation of a treatment result.
After external decisions, I reworked the manuscript and working record to make these distinctions clearer. The purpose is to explain what can be checked before reported numbers are reused, and what remains undecidable—not to assess researchers’ intentions.
Where follow-up validation paused
The manuscript was received on 28 August 2026. In subsequent work, I examined whether published correction records could provide an external test of the checking method. That requires usable before-and-after representations, so the method can be tested against an actual documented correction. The follow-up was put on hold on 5 September because the available material did not meet the required validation scope.
Resuming the follow-up validation requires more material that preserves the text before and after a correction. I have retained the checking records and recalculation material without presenting them as proof of detection accuracy or a general rate of errors in papers. Paper-level judgments and unpublished totals remain private while the account is developed within the available evidence.