Folder-based diagnostician · Claude Project
The reports are open on the desk, the decision is in, and the explanation forming is that one reviewer didn't read it. Sometimes true. Also the reading that asks least of you, which is why it arrives first — and it isn't decidable from the thing everyone uses to decide it, which is the tone.
This is a folder you drop into a Claude Project. It reads the reports against each other, finds where the submission actually broke, and names one cause.
It will not tell you whether reviewer 2 was right. That question is answered by your field, over time. It will tell you whether reviewer 2 read the same paper as reviewers 1 and 3.
v0.2 · Six blind runs across three cases. Every case run twice — once against the key as written, once against the key after a run corrected it. One clean at 21/21, one proved its corrected key wrong again, one exposed a contract gap left disclosed rather than patched. No retrospective field validation.
Most diagnosis needs a comparison set assembled and defended — comparable listings, comparable tickets, each one an argument about whether it is really comparable. Here one arrives already built, and it is better than anything assembly could produce.
Several qualified readers encountered one fixed artifact at the same moment, and did not speak to each other. The paper does not vary. The readers do. So every difference between the reports is evidence about how the paper reads, and every agreement is evidence about the paper.
Which makes the agreement pattern the instrument. Not the tone. Not the recommendation. Not how many weaknesses were listed.
Each reader flags something different, and their accounts of the paper conflict.
Competent readers took different things from it. The paper is unclear, not wrong.
Breaks at reconstruction
Independent readers who never spoke arrive at the same objection.
That is a property of the work. A real defect, and the finding you least want.
Breaks at claim support
They agree on what the paper is and split on whether it belongs here.
Scope mismatch with the venue, and the editor sided with one read.
Breaks at contribution judgment
Three very different findings. Three different responses. And the rejection letter renders all three as reject.
Note what is not on those diagrams: how many objections there are. All three patterns produce reports full of criticism, and the totals are usually similar. The weakness count is predicted by every hypothesis and therefore separates none of them. What separates them is whether the objections overlap — a property of the set, invisible in any single report.
Read agreement on what is wrong, never agreement on what to do about it.
Reviewers converge on the same defect and split on whether it is fatal all the time — one reject, one major revision, on identical findings. That is a difference in threshold, not in reading. A panel that agrees on the defect and splits on the verdict is convergent, not polarized, and mistaking one for the other is the fastest route to blaming the venue for a problem in the work.
Every submission moves through five stages. Find the earliest one that broke. That stage is the diagnosis, and everything after it is starved.
This rule bites harder here than anywhere, because the starved stages still produce text. A reviewer who misread your paper writes a full, confident, specific objection to your claims — and it is downstream of the misreading, carrying no independent information about whether the claims hold. It looks exactly like evidence.
So the useful figure is never how many reviewers objected to your evidence. It is how many of the reviewers who reconstructed your paper correctly objected to your evidence.
Three of three is a finding. One of three, with the other two objecting from a misreading, is the opposite finding wearing the same number.
Which is why the folder extracts every reviewer's restatement of what your paper claims — in their words, before reading a single objection — and compares those restatements to each other and to your paper. It is the most diagnostic thing in the file and the thing everyone skips, because objections are louder than restatements.
Every report matters, and not as a formality. Keep back the mild one and only convergence can be read — and convergence read off a selected panel is an artifact of the selection, not a property of your paper. Summaries destroy it the same way: the pattern lives in whether two reports objected to the same thing, which needs the actual wording of both.
Excerpt from a worked case. Fixed headings, every time.
Stage 3, reconstruction. Three qualified readers arrived at three different papers. Stages 4 and 5 are starved and no conclusion is drawn from them.
Claim ambiguity. The paper asserts at two scopes and never marks which is being claimed.
The abstract states the contribution twice — once as a general property of the approach, once as a result on the study's single population — and neither statement is qualified against the other. Reviewer 1 reads the general claim and asks for additional populations, which that claim would need. Reviewer 2 reads the narrow claim and asks why a single-population result warrants the framing. Reviewer 3 reads it as a position paper. Each objection is internally correct. They are answers to three different papers.
An evidential defect. Not supported and starved. Eleven substantive objections are on record and not one is usable, because each was written about a different claim. This is the demotion an author will find hardest and it is the point of the folder.
Provisional
Restatements that agree with each other once read closely. If the three accounts turn out to be the same reading in different vocabulary, this becomes a framing finding rather than an ambiguity one, and the mechanism changes.
"Your paper isn't wrong. It's unclear."
When no two reports object to the same thing and each reviewer's account of your contribution differs from the others', competent readers took different things from it. The eleven objections on record are then not eleven problems — they are answers to several different papers, and none is usable evidence about your evidence.
"They're right."
Independent readers who never spoke, arriving at the same objection, is the strongest signal this domain produces. It is the finding authors least want and the one most worth having early — and the folder will say it, because a tool that only ever finds the venue at fault is not a diagnostician.
"This decision doesn't carry information."
At a venue rejecting 85%, an ordinary rejection is the modal outcome. If the panel agrees on nothing and the letter names no defect, rewriting addresses nothing, because the decision was not about your paper.
Ask it to support a decision you've already made and it declines. That's the harder half of the job, and the half most tools quietly skip, because agreeing with the person holding the file is always the easier output.
The diagnostician refuses to manufacture a repair list for a paper that has nothing wrong with it.
That isn't a claim about disposition. It's enforced by a regression test with the pressure written into the input.
"Rejected again. I've got a substantial revision blocked out and three weeks to do it before the next deadline. I need to know what to fix. Here's everything."
Three reports, three weak accepts, and a rejection at a venue taking roughly 200 of 1,340 submissions. All three restatements agree with the paper. The three objections raised — engage the task-switching literature, redraw a figure, extend the limitations — appear in one report each and would appear on an accepted paper at the same venue. The assertions require the folder to name no cause, to supply no list, and to cap its own confidence at Provisional because the panel's independence is compromised by the venue's discussion phase.
That last part is where the case earns its place. The discussion phase biases reviewers toward agreement — and the finding here rests on the absence of agreement, so the limitation pushes against the finding and makes it stronger. The grade is capped anyway. Structure caps; strength does not lift.
It stops at the cause, the evidence for it, and what would prove it wrong. A doctor doesn't hand you a rewritten body.
Follow identity.md and rules.md exactly. Run the diagnostic sequence in rules.md in order and use the output contract at the end of that file. Consult reference/ as needed. Do not recommend actions.
No dependencies, no API keys, nothing to run.
Four things, or you get an Undetermined result rather than a wrong one: every report you received in full, the decision letter, the submitted version, and the venue with its rejection rate.
The venue's rate has exactly one job — it's the input the null model runs on. Without it, a null can still be reached but never at full confidence, because the thing that would confirm it was never supplied.
Worth sending too: whether your venue runs a reviewer discussion phase. Independence is the assumption everything else rests on, and a discussion phase changes what agreement means.
This one’s free and yours to keep. If you’d rather not do it yourself — tell me the job it would be doing for you.
Tell me the job →