Let me start with what works against me, and it is a fair amount: you already have a system. The editor’s letter, the response-to-reviewers table in a separate Word file, the comments stuck in the margin, and the discipline of crossing lines off as you fix them. It works. Plenty of people publish that way for an entire career and need nothing else, and this page is not going to pretend that method is broken.
The thing is that between the «major revisions» and the day you actually sit down to correct, there is teaching, fieldwork for another project and two weeks of August. When you come back, the response table is still there, intact and orderly, but you no longer know which of those twenty-odd objections you patched on the fly the first day, which one you rewrote from scratch, and which one you told yourself you would fix «at the end» and never did. Reconstructing that by hand, paragraph by paragraph, is redoing work you already did once. And the worst part is that you believe you did it: the feeling of having fixed something is just as vivid whether you fixed it or not. This tool is not going to tell you whether you fixed it either; what it does is show you where the text moved and where it did not, which is a question that has a checkable answer.
This is a revision case file. It keeps the objections from your previous report and, when you come back with the corrected version, it places each one on your new text: whether you have touched the paragraph it was anchored to, or whether the sentence it quoted is still where it was. What it never tells you is that one of them is resolved, and the whole reason is further down, because it is the argument I am proudest of and also the one that sells worst.
The response-to-reviewers table does one thing well and does not do the other. It keeps what you were asked for —reviewer 2’s comment, your reply, the page where you changed it— and that is genuinely good: it is a legitimate method, free, decades old, and the one the journal will ask you for anyway. What it does not do is anchor itself to a specific place in the manuscript. The table says «page 14, third paragraph», and page 14 has moved three times since.
The gap opens in the time between, not in the method. Weeks pass between the editor’s letter and the day you really sit down to correct, and you do not come back as the same reader: you come back as somebody who remembers the tone of the letter and not the detail of the objections. The journal’s reviewer, on the other hand, will come back with the report in front of them. That asymmetry is the whole problem.
And it was our problem too, not only yours. In the project’s own database, as of 31 August 2026, 31 % of the people who use the reviewer run more than one report, and 37 people have come back on a different day to request a new one. Until now, every one of those times started blank: the tool reread the manuscript as if it had never said anything before. The case file exists because that gap was ours before it was yours.
When round 1 finishes, your report carries a case number shaped like this: Q1-4H7K-9TXQ. It is inside the annotated Word file. With it you open the case file weeks later without having to hunt for anything else. It will ask for the same email address you used for the report; any address works, a personal one included.
There are two doors, and there are two on purpose. The number lives 30 days, which is exactly how long the privacy policy promises we keep the report: not one day more, because saying we store it for less and storing it for longer would be lying in the fine print. The Word file of your review never expires, because it carries the signed case file inside the file itself; if four months have passed and the number no longer works, you upload that .docx and carry on just the same. And if you opened the case file with that document and it happens to be your new version too, it cross-checks against it directly.
Then we ask you, without requiring it, for the version with Word’s track changes on. It is usually not an extra chore: most journals ask for the marked version alongside the clean one, so it is probably a file you already have on your desk. With the marks we look only at the paragraphs you touched plus the ones each objection quoted, and the check comes out more precise and cheaper. Without the marks it still works: we compare by the sentence each objection quoted.
The cross-check is computed in your browser. For that comparison the document does not even leave your computer, neither this round’s nor the previous one’s: it is not uploaded, not stored and not read on a server. From the case file we keep the objections —which are our own text— and the sentence each one quoted. Not the file, not the full text, not the title.
The full round is a different thing and it belongs here rather than in the small print: there the manuscript is read on a server, and while you pay it waits encrypted and tied to your payment for 24 hours at most, so it is not lost if you close the browser. It is deleted as soon as the report is generated; it is not stored. That is how the privacy policy words it, and the policy is the wording that governs: if this page and that page disagreed, that one would be right.
Prices, here for the same reason. Cross-checking against the previous round costs nothing. The full round on your new version —reading the whole manuscript again with two simulated reviewers and a simulated editor, with what the previous round said in front of it— is charged like any other: €12, one-off. If the second round ever becomes included with a previous Professional report, it will say so here and on the screen; today it is not, and I would rather you read that on the landing page than at the checkout.
Privacy policy: what is stored and for how long
When you cross-check your new version against the previous round, every objection from that round comes back with one of these two labels, in the same words you will see on screen:
It is «new»: this was not in the previous round. It cannot come out of the cross-check, and the reason is mechanical rather than commercial. The cross-check compares text: it takes the objections it had already stored and looks at where each one has ended up in your new version. To find something that was not there before you have to read the whole manuscript again, and that is the full round, which does cost money. In the free cross-check that box of the count reads zero, and I will now tell you why it is there anyway.
Comparing the quoted sentence is also not a character-by-character match. Before comparing we fold case, accents, curly quotes, dashes and runs of whitespace, because otherwise swapping straight quotes for curly ones when you move between editors would hand you half a dozen false «you touched this». So «still standing» means the sentence is still the same sentence, not that not a comma has moved. And if you upload the track-changes version, what decides it is whether the paragraph carries marks.
Because we cannot measure it yet, and I would rather tell you here than pretend otherwise.
Marking an objection as resolved requires a rule of equivalence: deciding that the objection from round one and the new paragraph in round two are about the same thing, and that what it asked for is now there. That is not settled by looking at whether the text changed —changing text is easy and fixes nothing by itself. It is settled with a corpus of manuscripts labeled twice over, with two people deciding separately whether each objection was covered, and with the agreement between those two people published so that you can look at it yourself. That corpus does not exist yet. Until it does, any «resolved» I showed you would be a model’s opinion dressed up as data.
And this particular error has a property that makes it worse than the others: it does not surface today. If I show you a «still standing» and I am wrong, you open the paragraph, see that you did fix it, and get on with your life in thirty seconds. If I show you a «resolved» and I am wrong, you close the case file at ease, send the manuscript, and find out weeks later, in the journal’s email, that the point you counted as closed was exactly where it had always been. An error that puts you at ease is far more expensive than an error that annoys you.
So the most we tell you is where you have touched the text and where you have not. Whether you have resolved an objection or merely moved it somewhere else is yours to judge, because you are the one who knows what the reviewer meant. When the validation exists —the corpus labeled twice over and the agreement between labellers published, in that order— the fourth state will appear, and this page will report it with the agreement figure attached. Until then, that last step is yours, which is the only place it is currently taken well.
In the count you will also see a «not checked» box, and it is worth knowing what it is before you read it as good news: it is not a new state, it is the same «still standing» telling you how far it could look. It happens for two reasons. One, the objection was about something missing —«you do not report the effect size»— and quoted no sentence, so there is nothing to compare. Two, the sentence it quoted is too short to locate unambiguously: a «p < .001» appears half a dozen times in a manuscript and makes a useless anchor. In both cases the objection is still standing; what we cannot do is tell you whether the text moved.
The «new» label looks like the least important of the three and it is the one with the data behind it.
In the project’s peer-review corpus —5,439 pairs harvested from F1000 and PCI, snapshot of 30 August 2026— the measurement was run on 265 manuscripts with more than one version. Between the first version and the last, objections drop from 13.92 to 3.02 on average. That is the pretty number and it is true, but on its own, without what comes next, it is worth nothing.
The caveat, which is what separates this figure from a brochure number. A manuscript with several versions is, by definition, one that was asked to revise: it starts off worse than average, so part of that drop is the floor it came from rather than the merit of the revision. And those two platforms are publish-then-review, not the circuit of a traditional journal with two closed rounds: their rounds do not look like your Q1’s. If somebody shows you a drop like that without telling you where the sample comes from, be suspicious; this page included.
And the figure that really justifies the screen: of those 265 manuscripts, in 28 the objections went up after revising. One in nine gets worse while being fixed. You rewrite the discussion to satisfy reviewer 2 and the new paragraph opens a flank that was not there before; you trim methods to save space and leave the exclusion criterion out. That is why the case file is not a counter that goes down, and why there is a label that says «this was not here before» —the one that fills up in the full round, when the manuscript is read again from end to end.
Starting where I lose. Paperpal Prime costs $139 a year and includes unlimited submission checks; without a subscription its check is free and you only pay $29 if you want to download the processed manuscript. If what you need is to run the text through many times, that is cheaper per pass than buying single rounds from me, and there is no arguing with it: it is arithmetic.
And continuing with where this page was wrong until today. The first version said the market sells repetition rather than memory, and that the market’s general answer to «what does the second pass know about the first» was «nothing». That is false, and I found out by going back to the sources before publishing. Manusights charges $39 per report, sells 3-packs and 5-packs at 24-30 % off —with the explicit argument that most manuscripts go through two or three rounds— and its FAQ answers that yes, it remembers: it says that when you re-upload a revised version, the engine reviews it in light of your prior round and tells you what you addressed, what still shows up, and what is new.
So I am not the only one who worked out that a manuscript goes through more than one round, and claiming the idea would be shabby. The difference sits one box further along, and it is exactly the one this page has spent five sections on. They tell you what you addressed. I do not, and it is not decorative caution: deciding that an objection has been addressed means applying a rule of equivalence, and a rule of equivalence with no doubly labeled corpus and no published agreement between labellers is not a measurement, it is a verdict. I have not seen that number published by any AI reviewer on the market, and I do not have it either: which is why that state does not exist here, rather than existing without it.
As for the rest of the market, only what the source holds up and nothing more: Writefull Revise hands you the revised document with track changes and is sold to publishers and institutions only, not to the individual author. And about the automatic review systems that have actually been studied, the paper that introduced the FirstPass dataset and model in June 2026 —3,668 complete peer-review dialogues from Nature Communications— says in its introduction, word for word: "Existing systems generate reviews in one shot, blind to the author-response iterations where scientific claims are actually stress-tested."
So the sentence that sums the section up is not the one I wanted to write, but the one that holds: against somebody offering you unlimited for $139 a year, boasting that you can repeat it is competing on their ground and losing; and against somebody who already claims to remember, boasting that I remember is not enough either. What is left, and it is the only thing I am offering, is that here you fill in the last box yourself for as long as nobody can show the number it would take to fill it in for you.
Paperpal — pricing · Manusights — AI review and FAQ · Writefull Revise · FirstPass (arXiv:2606.20769)
Yes. Upload the Word file of your review: it carries the signed case file inside the file itself and does not expire. The number, by contrast, lives 30 days, which is how long we keep the report.
No. You need the same email address you used for the report, whatever it is: that is how the case file is found. If you signed in with your university, hospital or center address, your center counts on the adoption map, but it is not a condition.
No, but it helps. With the marks we look only at the paragraphs you touched plus the ones each objection quoted, and the check is more precise. Without the marks we compare by the sentence each objection quoted. And it is usually a file you already have: most journals ask for the marked version alongside the clean one.
In the cross-check, no: it is computed in your browser and the document does not leave your computer, neither this round’s nor the previous one’s. From the case file we store the objections, which are our own text, and the sentence each one quoted. In the Professional round the manuscript is read on a server, and while you pay it waits encrypted and tied to your payment for 24 hours at most; it is deleted as soon as the report is generated. The privacy policy spells this out.
Because we cannot measure it yet. Deciding that two objections are the same one, or that one no longer applies, is a rule of equivalence that has not been validated, and it needs a corpus labeled twice over with the agreement between labellers published before we show anything of the sort. A wrong «resolved» would leave you at ease about something that is still there, and that does not surface today: it surfaces weeks later. When the validation exists, the fourth state will appear with its agreement figure beside it.
Cross-checking against the previous round costs nothing, and you can repeat it with as many documents as you like. The full round on your new version is charged like any other: €12, one-off.
No, and that is the whole difference. Running it again is a first round all over: the new report knows nothing about the previous one. The case file brings back the previous round’s objections and places them on your new text; the full round, on top of that, reads the manuscript again with two simulated reviewers and a simulated editor, with what the previous round said in front of it.