Search for a command to run...
A consensus row in our database describes one supplement-and-outcome pair, and it carries two counts that look almost identical. One is the number of studies. The other is the number of claims with a recognized direction — increase, decrease, or neutral. They sit on the same row. They are different numbers. We did this on purpose, and the reason is the kind of small decision that decides whether a page adds up.
Here is the problem. Not every claim extracted from a paper has a clean direction. A study might report that a supplement affected an outcome without specifying whether the effect was an increase or a decrease, or the text might be ambiguous enough that we cannot responsibly bucket it. Those claims are real — they came from a real study — but they cannot contribute to a directional bar chart, because a bar chart needs to know which way the arrow points.
So you have a choice, and it is more consequential than it looks.
You could count only the directional claims. Then your bar chart of increase versus decrease versus neutral is honest, and the counts add up to the bars. But your "number of studies" header now disagrees with the bars, because studies that only contributed ambiguous claims vanish from the directional count while still existing in the literature. A user sees twelve studies listed but a bar chart that sums to nine. Something looks wrong, even though nothing is.
Or you could count every study for the header, and count only directional claims for the bars. Now the header and the study list agree — every study you list contributes to the number — and the bars are honest about the directional split. But the two numbers on the page do not match, because they are counting different things.
We chose the second, deliberately.
The row carries study_count — the number of distinct studies in the bucket, counting everything, including studies whose claims had no recognizable direction. And it carries the directional counts — increase, decrease, neutral — which sum to something less than or equal to the study count, because ambiguous claims are excluded from the split but not from the existence.
The header shows the study count. The bar chart shows the directional split. They are different denominators, on the same row, and that is correct.
The temptation is to make them match, because matching numbers feel safe and mismatching numbers feel like a bug. But matching them requires lying in one direction or the other — either inflating the directional bars by pretending ambiguous claims have a direction, or deflating the study count by pretending ambiguous studies do not exist. Both are worse than an honest mismatch that the reader can understand.
When two measurements on the same object disagree, the instinct is to find the bug. Sometimes the disagreement is the truth, and forcing agreement is where the lie enters.
Our counts do not match because the world they describe does not match: some studies contribute cleanly to a direction, and some do not. Carrying both numbers, clearly labeled, lets each one be honest. The user sees twelve studies and a nine-claim directional split and understands, correctly, that three of those studies did not point a clean way. That is a more faithful picture than twelve and twelve, which would be tidy and false.
The cleanest page is not the one where all the numbers agree. It is the one where every number means exactly what it says.
