Numbers that don't add up: catching numerical errors in your manuscript
Why a correct result can still be reported with the wrong number, and a method to catch it before a reviewer does.
Most numerical errors in papers are not errors in the analysis. The experiment is fine; the paper reports it inconsistently. A result is re-run and the table is updated, but the abstract still quotes the old value. A total is written down before one subgroup is added. A figure label is copied from the previous version of the figure.
To a reviewer, though, two different numbers for the same quantity look the same whatever the cause: one of them is wrong, and they can't tell which. That doubt spreads to the rest of the results.
Four ways it happens
1. The same result, quoted in several places
A headline number typically appears in the abstract, the introduction, the contributions, the results and the conclusion. Only one of those is next to the data.
From a machine-learning paper
The contributions list reports peak memory savings of 22.8% and 14.9%. The results section, measuring the same models against the same baseline, reports 23.8% and 15.6%.
Fix: use the results-section numbers in the contributions, where most readers will see them.
2. Totals written before the parts changed
Counts, sample sizes and percentages are summed once and rarely re-summed.
From a medical-imaging paper
The text names four organs, then gives counts for three of them and “a total of 281”. The three counts already sum to 321 (131 + 64 + 126). The fourth count was missing, and 281 was in fact the size of that missing set.
Fix: add the missing count and recompute the total, or drop the word “total”.
3. Numbers inside figures
Figures are made in a plotting tool, outside the text, and proofreading tends to skip their labels.
From an economics working paper
A figure is labelled “Simulations Run: 2 × 5 × 8 = 405”. Two pages earlier the prose has it right: 2 × 5 × 8 = 80. The figure label had been copied from an earlier figure.
Fix: change the label to 2 × 5 × 8 = 80.
4. Text and table on the same page that disagree
Even when the table sits right next to the sentence describing it, the two can drift apart if the table is regenerated and the sentence isn't.
From a genomics paper
The results text says two language models reach 0.74 and beat an expert baseline at 0.72. In the table on the same page, the baseline is 0.74 and the two models score 0.78 and 0.79.
Fix: update the sentence to the numbers in the table.
A method that works
- List your load-bearing numbers. Headline results, sample sizes, dataset sizes, hyperparameters that matter. Ten to twenty numbers for a typical paper.
- Search the final PDF for each one and note every page it appears on, in the text, tables, figure labels and captions. Also search for the rounded and unrounded forms (23.8, 23.79).
- Recompute every total, average and percentage from its parts, including the ones printed inside figures.
- Check comparisons in both directions. “A beats B” needs A's number to be larger in the table, for every case the sentence claims.
- Fix the rounding rule once (for example, one decimal for percentages) and apply it everywhere, including tables.
- Repeat steps 2 to 4 after the last re-run. This is the step that usually gets skipped, and the one that matters most.
Where the method runs out
A manual ledger catches repeated numbers you thought to list. It misses the ones you didn't: a count in a figure legend, a percentage in a footnote, a value that appears once in the text and once inside a table image. That is the part worth automating. DeskReady reads every page, including figures and tables, and reports each mismatch with both places quoted and highlighted.
For the wider final-read list, see the pre-submission checklist. For sentences that disagree with tables specifically, see when your text and tables disagree.
DeskReady runs these checks on your PDF in a few minutes and shows every finding highlighted on your own page. Your first check is free and needs no account.
Check your paper