Reviewer 2’s note was polite but firm: the authors should engage with the original trial instead of leaning on secondary summaries. Fine, I thought. I’ll add the citation. The trouble arrived an hour later. The trial was a 1998 Japanese clinical study cited in every English-language review of the intervention, and no one in my department had actually opened it. WisPaper, an AI academic agent, was how I eventually read it, but that came later.
Getting the PDF took a week of interlibrary requests. Opening it solved nothing. The full text was Japanese, and the English abstract reported results while skipping the methods entirely. Every review I had read paraphrased that same abstract, so I was holding the primary source with no way to read it.
The Citation Everyone Skips
A citation trail runs like a game of telephone. Each review borrows the sentence before it, and after twenty years the trial’s methods have flattened into a secondhand summary. I gave Deep Search the title, the year, and the treatment name, and the agent confirmed the record and pulled the full text my usual databases had buried behind broken links and paywalls. That told me something: two decades of citations from people who never saw the source.
What Hides in the Figure Captions of a 1998 Paper
With the full text in front of me, the diagrams were at least legible. Figure 2’s three-line Japanese caption, beneath a line drawing of the dosing schedule, described an interval every English summary had gotten wrong. The reviews said the drug was given weekly. The caption said twice weekly, with a washout period none of them mentioned. The key protocol detail lived only in that caption, a place most citing authors had never looked, so the field had spent decades repeating an assumption instead of the method.
Reading a Language You Don’t Read
The immersive reader’s full-text bilingual translation handled my Chinese and English sources without complaint, and its easy-reading mode untangled the dense layout so each figure sat next to its own caption. But the Japanese primary text needed a different lever, and that is where the case finally turned.
Asking Questions Only the Original Authors Could Answer
I stopped trying to read the paper like a fluent speaker and started asking it questions, the ones every review should have asked. WisPaper’s Scholar QA answered from the document and attached a chain of evidence to each claim, pointing back to the exact figure caption and the methods line. I could verify every answer against a specific location instead of trusting a paraphrase.
Bringing the Missing Protocol Into Your Review
I rewrote the methods paragraph with the twice-weekly schedule and the washout window, cited against the original figure. The reviewer signed off. What changed was smaller than it sounds: a review that states the method rather than the assumption it inherited. An AI academic agent that can read a paper you cannot, and show you where each claim comes from, turns a citation ritual back into an encounter with the source.
