When a Citation Becomes a Citation of a Citation
Every English-language review in your subfield leans on this trial. It anchors a dosing claim, or a safety claim, that no one can quite reconstruct anymore. Trace the footnote back far enough and the same sentence turns up in twenty papers, each one pointing at the next. Open any of them, and you find the same half-sentence, reworded just enough to avoid looking copied. The original is never the source. The field has been quoting a quotation — the same gap that shows up whenever a widely cited statistic loses the context that shaped it.
This isn’t a rare accident. Citation-chain studies have flagged the pattern for years: a claim gets simplified once, then simplified again by the next author who cites the simplification instead of the source. By the fifth generation, the sentence has drifted far enough from the original data that nobody citing it could tell you where the number actually came from. Nobody meant to do it. The abstract was in English, the summary sounded solid, and the paper moved to the next citation slot.
What a 1998 Figure Caption Can Hide From an Entire Field
The protocol detail that would change your methods section sits only in the caption of a figure almost no citing author has actually seen. It describes a washout period between the two treatment arms — a detail no English summary mentions, anywhere. Reviews across the field have repeated an assumption instead of the method itself. One caption. That’s the whole gap between citing a result and understanding it.
Figure captions carry this kind of weight more often than researchers expect. A caption is where an author writes the thing that didn’t fit cleanly into the prose — a control condition, a sample-size adjustment, a timing detail nobody thought to spell out twice. Abstracts get translated for indexing. Captions almost never do. So the detail that would actually change how you design your own trial sits in the one part of the paper that machine indexing skips and human reviewers, working from an English summary, never reach.
Reading the Paper Instead of a Digest of It
This is where WisPaper does something plain and useful. Its full-text bilingual translation opens the methods and the figure captions side by side, so you move through the details yourself instead of trusting someone else’s summary. You read the page. Not a paraphrase of it.
That distinction matters more than it sounds like it should. A machine-translated abstract can tell you what a paper concluded. It can’t tell you what a paper actually measured, what it controlled for, or what got mentioned once in a caption and nowhere else. Conclusions survive translation. Methodology, the part that determines whether your own citation is honest, tends to get lost in exactly the sections that summaries skip.
Interrogating a Paper the Way You’d Interrogate Its Authors
You can also ask the paper direct questions, the way you’d question its authors if they were still reachable. Ask where the washout interval is specified, and the answer comes back with a chain of evidence pointing to the exact sentence and page. Not a third-hand rendering. A location you can check yourself.
That’s a different relationship to a source than skimming a translated abstract and trusting it. It puts the paper itself back in the loop, instead of the twenty papers that have been citing each other about it. And it scales past one trial — the same approach works whether the buried detail is in Japanese, German, Korean, or any language a reviewer happens not to read.
Putting the Missing Protocol Back Into the Record
You rewrite the paragraph in your own methods section. For the first time, it says what the trial actually did. A reader following your footnote now lands on the figure caption, not a dead loop of secondary citations.
Multiply this by every subfield where a foundational study happened to be published somewhere other than an English-language journal, and the scale of the problem stops looking like an edge case. It looks like a structural gap in how research actually gets read. That’s the drift an AI academic agent fixes — a field slowly separating from what its sources actually say, one untranslated caption at a time.
