Mindmap and Review a Paper

Reviewers check whether your sources say what you claim they say, and whether your conclusions follow from your claims. This guide runs both checks before you submit. You turn a paper into a mindmap of its claims, every claim is tested against the papers in your library, and you get a verdict under each one. Then the argument across the claims gets a verdict of its own, all in one go. Think of it as an AI peer review of your paper, with every verdict pinned to the page that settles it. It works the same on someone else's paper.

The screenshots come from a real run. We wrote a two-page draft review on social media and adolescent wellbeing that cites eleven papers, ten of them in the library, and planted seven mistakes in it: an inflated number, two reversed findings, an association stated as a major risk, a finding pinned on the wrong paper, a source missing from the library, and a conclusion that goes further than the evidence. The review caught all seven.

Before you start

• Put the papers you cite in one library. The review can only find support in papers your library holds. Upload the PDFs or send a Zotero collection, see the Quickstart and Zotero Integration.
• Add your draft to the same library as a PDF and wait until it is indexed. You will not cite it, the agent only reads it to list its claims.
• Use the Starter plan or higher for the checks. Mapping the claims alone works on every plan.

Step 1: Map and review the paper

1. Open the project's mindmaps and click Mindmap and Review a Paper, under + New Mindmap.
2. Click the paper. The list shows every indexed paper in the libraries attached to the project, and the filter box narrows it by title or author. When a paper has several files, such as a main text and a supplement, each file is listed by its file name, and you map one file at a time.
3. Choose what happens after mapping. Both boxes are ticked by default:
   • Check support gives a verdict and evidence for every claim.
   • Check the reasoning tests whether the conclusions follow from the claims.
   
   The panel shows the paper's page count and the estimated credits, which update as you tick or untick the boxes.
4. Click Run.

A new mindmap opens right away, named Claims in followed by the paper's title. A progress note shows each step: mapping the claims, checking them against your library with a running count, and checking the reasoning. Everything runs on our servers, so it keeps going if you close the tab, and the results are there when you come back. Our draft gave 19 claims, and the full review took under 2 minutes.

If mapping fails, the empty mindmap is removed and you pay only for what was used. If a check cannot finish, the claims stay and a note says which check stopped.

What the mindmap holds

The mindmap has one branch per section of the paper and one node per claim.

The agent lists the sentences a reviewer would check: findings and results with their numbers, causal, comparative, or general statements, statements about other studies such as "Smith (2019) showed ...", and the authors' interpretations and conclusions. Descriptions of what the paper will do, method details, definitions, captions, the reference list, and acknowledgements are left out.

• One claim per sentence, the way a reviewer reads. A sentence is split only when its parts cite different sources.
• Each claim stands alone. "This effect" becomes the effect it refers to.
• The paper's own certainty and scope are kept. "May", "suggests", and the population studied are never strengthened or dropped.
• In-text references are kept as written, such as "(Collis and Eggers, 2022)", so the review knows which source to look inside.
• Claims carry no citations, on purpose. A citation would point back at the paper itself, and every claim would look supported.

Validation groups

Several claims often rest on the same evidence: two sentences that cite the same trial, or an abstract claim and the discussion claim it summarizes. So each claim carries a note naming its validation group, such as "Restriction experiments and recommendations (Collis and Eggers 2022, Walsh et al. 2024)". The claims of one group are checked together: one agent looks up the shared evidence once, then writes a separate verdict under each claim. That is cheaper than checking claims one by one, and a conclusion is judged against the same passages as the findings it rests on.

To see a claim's group, select the claim and open Properties in the right sidebar, under Notes.

Your paper is left out

The mindmap remembers which paper it came from. When the claims are checked, that paper and all of its files are left out of the library search. Your draft repeats every one of its own claims, so it would otherwise find itself first and vouch for itself. The agent can still read the paper to understand a claim in context, but the paper never counts as evidence.

Step 2: Read the verdicts

With Check support ticked, each claim now has a Library check verdict underneath it.

To reach a verdict, an agent works the way a reviewer does. It checks whether each paper the claims cite is in your library and searches inside it, searches all of the project's libraries for passages that back the claims and for passages that would contradict or narrow them, and reads the whole page when a passage is close but not decisive. It judges only from passages it found in your libraries, never from what the model may remember.

The verdict node is coloured from red to green by its score, so a section with problems stands out before you open anything. Not found in your library does not mean the claim is false. It means your library cannot back it, and the cited paper may simply be missing.

Here is what the review said about the planted mistakes:

None of the accurate claims came back Contradicted or Not found. They came back Supported or Partly supported.

Step 3: Open a verdict and fix the claim

Click the arrow on a verdict to expand it.

Your claim
└── Library check: Contradicted (8/100)
    ├── Supporting evidence
    ├── Contradicting evidence
    ├── Issues
    └── Suggested rewording

• Supporting evidence and Contradicting evidence list the passages, each with its citation. Click a citation to open the page in the reader and read the sentence yourself.
• Issues says what is wrong in plain terms: an overstatement of scope or certainty, a finding attributed to the wrong paper (the issue names the paper that does back it), a cited paper missing from your library, or a qualifier the passages state and the claim leaves out.
• Suggested rewording gives a version the passages support, with its citations attached.

Contradicting evidence, Issues, and Suggested rewording appear only when there is something to show. To fix the draft, copy the rewording into your manuscript or write your own. To cite the right passage in the mindmap, drag its citation onto your claim, then press F3 to run AI Verification, which checks the new wording against the citations it now carries.

Step 4: Read the reasoning verdict

A verdict per claim looks at one claim at a time. A reviewer also asks whether the claims, taken together, carry your conclusions, and a paper can pass every claim and still conclude more than its claims allow. With Check the reasoning ticked, the Reasoning check node under the root of the mindmap answers that.

The agent receives every claim with its validation group and its verdict, and reads the groups as your lines of evidence. It finds the main conclusions and the claims each one rests on, reads the pages of your draft where the conclusions are argued, and searches your library for findings your conclusions leave out. A conclusion that rests on a Contradicted claim inherits the problem.

The root of the mindmap
└── Reasoning check: Weak (38/100)
    ├── a short summary of the main conclusion and whether the claims carry it
    ├── Argument
    │   └── one node per main conclusion, with the claims it rests on
    └── Issues
        └── one node per issue, with the claims involved, passages, and a suggested fix

The issues it flags are the ones a referee raises: a conclusion that does not follow, an unstated premise, two claims that contradict each other, a conclusion resting on a weak claim, an overgeneralization from one sample or design, circular reasoning, and an alternative explanation in your library that the paper ignores. It lists the issues that change a conclusion, most serious first, and leaves out nitpicks.

On our draft it came back Weak at 38, with issues such as:

• The conclusion that limiting social media "reliably improves adolescent mental health" rests on the two trials whose descriptions came back Contradicted.
• Cross-sectional associations and a microsimulation do not show that limits improve mental health.
• A daily limit for every student goes further than any study in the library.

Each issue ends with a suggested fix, such as narrowing "reliably improves" to what the trials found.

Credits and plans

Mapping the claims is available on every plan and costs little: 0.2 credits for our two-page draft. The checks need the Starter plan or higher. The whole review settles as one Paper map line in your credits history. Our draft cost 5.5 credits in total: 0.2 for mapping, 4.8 for the claim verdicts, and 0.5 for the reasoning. Before it starts, the review holds an estimate of 0.1 credit per page for mapping, 3 credits per page for checking support, and 5 credits for the reasoning, then charges what the task used. A claim mindmap counts toward your plan's mindmap limit.

Let Claude run it

If you use Claude Code or another MCP client, the Agent Bayes MCP server can run the whole review for you and write up the result. Connect it, then ask:

Peer review my draft "Screen Limits and Adolescent Wellbeing" in my Social Media and Teen Wellbeing project. Map it with both checks. Give me the reasoning issues first, then the contradicted, contested, and not-found claims, each with the issue, the page that settles it, and a rewording.

Claude maps the draft with mindmap_map_paper, with check_library and check_reasoning turned on, waits for the review, then reads the results and reports. The verdicts land in the same mindmap, so you can open every page in Agent Bayes afterwards.

Tips

• Check the estimate on long papers. A thesis chapter can have a hundred claims, and the panel shows the estimated credits before you run.
• "Not found" is often a library gap. Before you rewrite a claim, check whether the cited paper is in the library.
• Review the revised draft. The checks run once, as part of mapping. After you fix the manuscript, add the new version to your library and map it again.
• Keep the claim mindmap. It is an ordinary mindmap with saved versions, so it records what you checked and what you changed.

What's next

• Peer Review a Paper: write a referee report on someone else's paper.
• AI Verification: check a claim's wording against the citations it carries.
• Outlining a Chapter or Paper: build the chapter from cited claims in the first place.