See how Thesisroom checks a thesis, and why it never writes one

Thesisroom is built for the defense, not for the writing. A tool that helps a candidate compose paragraphs ends up writing for them. A tool that checks a thesis after it is written holds the candidate to the standard a supervisor, examiner or committee will apply, and leaves every sentence theirs. Thesisroom does its work between the draft and the viva, once the candidate has written the text.

Thesisroom never produces thesis prose. It produces questions, margin comments and verification reports. The candidate writes the thesis. Thesisroom questions it.

Two kinds of work happen inside Thesisroom, and two different systems handle them. Facts about citations are looked up deterministically in published research databases such as Crossref and OpenAlex, drawing on ten registries, and every citation comment states its reason in plain words. There is no language model deciding whether a citation is real.

Arguments, questions and structural critique are handled by a language model that reads only the candidate's own thesis and its bibliography. It cannot invent sources. Between the two sits an audit step that confirms every claim the model makes points to a page in the thesis. Nothing reaches the candidate without a source.

The clearest case of this division is the Statistics check. It takes every significance test you reported in a sentence, recalculates the p-value from the test statistic and degrees of freedom you gave, and tells you where the reported figure cannot be right. It reads results written in your text, not results in tables. No model is involved at any point. You can verify any result it reports by hand with a statistical table, which is the standard the whole deterministic side is held to: if a supervisor cannot check it by hand, it does not belong on this side of the line.

The method is published. Thesisroom follows the recomputation approach of Nuijten, Hartgerink, van Assen, Epskamp and Wicherts (2016) in Behavior Research Methods, and the statcheck package by Epskamp and Nuijten, with one deliberate departure: a reported statistic is itself rounded, so Thesisroom computes the full range of p-values the rounding allows and reports a result only when the reported figure falls outside all of it. That is stricter than the published method. It reports fewer errors and never reports one that is not there.

The References check is the clearest case of the other half of that rule: a check does not have to compute anything, it has to be checkable. Citation Check has already parsed your bibliography into entries. References takes those entries and matches them against the citations in your own text, in both directions, and names the citations with no entry and the entries nothing cites. Nothing is computed and nothing is judged. Two lists are compared, and a supervisor can confirm any single result by looking the name up in your bibliography.

It also shows what the deterministic side owes you when it cannot answer. References does not run on a document with no reference list, on a bibliography Thesisroom could read only part of, or on a document whose citation style it could not tell apart, and each of those says which one happened and what went unchecked. A check that quietly reported nothing found in any of those situations would be worse than one that did not run, because a candidate would read the silence as a clean result.

The Structure check applies the same two rules to the document as an artifact. It reads the detected chapter list and every numbered figure and table reference written in your text, and reports where a sequence breaks, where a contents entry matches no detected heading, and where the word count exceeds a limit you set. Chapter 5 either follows chapter 4 or it does not, which is why no model is involved. And where a contents page cannot be read from the parsed text, it says the contents cannot be checked rather than reporting nothing found.

That division is the whole product: deterministic where determinism is possible, a language model only where reasoning is required, and the audit between them is what makes the model's output safe to read.

Citation Check draws on ten academic registries. Crossref and OpenAlex find the work. Then each registry is asked the question it can answer, and the result names every one that answered. Crossref returns the canonical DOI metadata. OpenAlex returns the relational graph: authors, institutions, related works. Semantic Scholar returns abstracts and the work's place in the citation network. PubMed confirms work curated by the National Library of Medicine.

Retraction status comes from the registry records: Crossref carries Retraction Watch's database on the original work, OpenAlex marks the work as retracted, and PubMed marks a retracted publication in its own article type. Where Retraction Watch holds the retraction, the result says so and links the notice. DOAJ is asked about the journal by its ISSN, and the result says when the journal is listed in the Directory of Open Access Journals. OpenAIRE and HAL report repository and preprint copies and Unpaywall locates open-access copies; they add context, and none of them verifies a source on its own. A work OpenAlex holds only as a repository copy is found, not verified, in the same way. A citation that fails any of these checks is marked on the page, with its reason in plain words and a link to the paper where one exists.

DataCite is asked about a DOI none of the registries above knows. It registers DOIs for datasets, software and theses held in repositories, and its record names the type of work, which the result shows in plain words.

Every question Thesisroom writes quotes the passage in your thesis that produced it, and names the page that passage sits on when you upload a PDF. Every margin comment is anchored to a sentence or paragraph. Every citation comment states its reason beside what you wrote, with a link to the paper where one exists.

The candidate can verify everything Thesisroom says. So can the supervisor. Thesisroom is not asking to be trusted on its word. It is asking to be checked.

Reference Watch and Paper Match keep working after a run finishes. Reference Watch re-checks your bibliography against the same ten registries on a weekly cycle, entirely deterministically, and raises an alert only when a source's status actually changes, with the registry that changed it. Nothing is added to your bibliography on its own.

Paper Match runs on demand. Upload a published paper and Thesisroom reads it against the claims already anchored in your thesis, classifies each relationship as supports, contradicts, complicates, or neutral, names any methodological gap it reveals, and checks the paper against the registries. The verdict is a classification. Whether to cite is your decision.

Blind Spots asks the opposite question: not what a paper you found means, but what the field holds that you have not cited. For each claim it counts the highly-cited work on that topic in OpenAlex, a published research database, and compares that count against your bibliography, giving you a ratio and the specific papers behind any shortfall. The ratio is arithmetic. Whether a gap matters for your argument is your judgment.

Argument Map reads your thesis against itself and stays open while you write. It maps every claim, defined term, and research question as a verbatim quote with its page, then tracks which definitions drift between chapters, which questions go unanswered, and which claims sit in tension. Searching that map consults nothing but your own stored words. It never tells you how to resolve a conflict; it shows you where one is.

Beyond the ten rooms, the workspace supports the daily research workflow: a report your supervisor can read before your meeting, a record of what they asked for and which draft it was raised against, a pack assembling both before the next meeting, a weekly list of what the record holds near your thinnest claims, a research journal nothing reads back to you, and a view of which chapters carry a publishable contribution.

Before any analysis runs, Thesisroom removes the student's name, email address, phone numbers, student ID numbers, title page, declaration and acknowledgements from the thesis. A supervisor's or institution's name can remain where it appears elsewhere in the text. No language model provider Thesisroom uses trains on uploaded theses, by its own terms. Students can delete a thesis or their whole account at any time. Thesisroom does not write, rewrite, or ghostwrite any part of a thesis.