For people who read the references
Read the paper with every citation checked on the page
Upload a PDF. Each citation is checked against the published record and marked where it sits: verified, differs, retracted or not found. Ask Archana, and every sentence of her answer shows its source. Then grow the paper into screened related work and a dataset where every cell opens its quote.
- Free: 1 paper a month, no credit card
- Your papers, notes and chats never train AI models
2.1 Reprogramming routes
Somatic cells can be returned to a pluripotent state by four transcription factors [9], verified, and the same factors reprogram adult human fibroblasts [12], differs. Later work reported pluripotency induced by a brief exposure to low pH [14], retracted, and reprogramming without any added transcription factors [17], not found.
Citation [14]
RetractedClaim · p. 4
“pluripotency induced by a brief exposure to low pH”
Reference
Obokata H, et al. Stimulus-triggered fate conversion of somatic cells into pluripotency. Nature. 2014;505:641–7.
Published record
Retracted by Nature, 2 July 2014.
The film
A second brain that checks itself
Transcript of “Citation Monster, the film”
Every researcher wants a second brain. Building one by hand is a second job. Now, you just drop in the paper. It reads every page and checks every paper it cites. Checks out. Wrong year. Retracted. And this one, nobody can find. Anything shaky is flagged before it goes in. Every fact is saved with its page. It pulls in every paper this one cites and checks those too. Then, the papers they cite. Explore it by topic, by section, or by who wrote it. Ask it anything. Every answer shows its source. Turn a shelf of papers into one table, every cell sourced. Your whole lab can add to it. A second brain that checks itself.
The tour · 2:47
See it on one paper
Follow a single paper outward: what it says, the forty references it stands on, the work behind those, the people, and what came after. Under three minutes, recorded in the app.
Transcript of “One paper in: a tour of Citation Monster”
Every paper is a doorway. Most of us read it, and leave. Let's walk through one instead. Attention Is All You Need. One paper in, and we'll come out knowing the literature behind it. Start with the paper itself. The page stays clean. Every citation is marked right where it's made, and J and K step through them. Everything it reports is already pulled out: every fact, every claim, each with its quote and page. And you can ask it anything. Is the French BLEU score consistent? No. The abstract says forty-one point eight, section six says forty-one point oh. Click the page, see the words. Ask what it doesn't cover, and it says so, instead of guessing. Now step back one ring. Forty references. Hit Re-check all, and every one is checked against the public record, live. Thirty-nine verified. One is cited with half its title missing. The record has the full title. Now you do too. Then the harder question: does each source actually back the sentence that cites it? Forty-six claims supported. None contradicted. And you can see where each one is used: which section leans on which source. Every reference has references of its own. LSTM, from 1997, stands on thirty-four more. One click. Turn them all on, and the graph goes two levels deep. The clusters are the schools of thought this paper grew out of. Recurrent networks. Optimization. Parsing. Vision. Prefer subjects? The same map, by topic. Behind every cluster are people. Click a name: Sepp Hochreiter. His work, his collaborators, where he shows up in your library. So far, we've looked back. Grow looks forward: who built on this paper, who tested it, who pushed back. You set the question and a spend cap. Four hundred papers found, the top ten ranked, for under a penny. Like this one: can a pure attention model learn images without a giant dataset? Archana suggests: include, with the quote that says why. You decide. In, out, or not sure. Anything you keep comes into your library, checked the same way. Any paper's results become a table, every cell tied to its quote. Export it all, with the sources attached. And everything you've touched stays connected, ready to ask. One paper in. The whole literature behind it, checked. Find the reference that doesn't hold up, at citationmonster.com.
Every citation is a chain
Claim, reference, source: checked link by link
A citation is only as good as its weakest link. Citation Monster checks that the reference exists as printed, then reads how the sentence uses it and whether the source says what the sentence says.
- Each reference is checked against the published record: verified, differs (with the corrected details), retracted (with the date and source) or not found.
- “Audit this paper” sorts every citing sentence into supports, contrasts, mentions or uses its method or data.
- Support is judged from the best source available: the cited paper itself if it is in your workspace, otherwise the earlier web check or its abstract. A supported answer always quotes the passage.
- When no passage turns up, it says “not found in sources checked”, never “false”.
- References of references open in place and feed the knowledge graph; author names open scholar pages.
Audit this paper
4 claims supported (1 partly), 3 not found in sources checked, 0 contradicted.
7 citing sentences judged.
How the paper cites them
- Supports3
- Contrasts1
- Mentions2
- Method/data1
Do the sources support the claims?
- Supported3
- Partly1
- Not found in sources checked3
- Contradicted0
Reference [9] · cited on p. 4
“Somatic cells can be returned to a pluripotent state by four transcription factors [9]”
From the source
“…induction of pluripotent stem cells from mouse embryonic or adult fibroblasts by introducing four factors…”
Takahashi & Yamanaka, Cell, 2006 · abstract
Ask Archana
Answers that show their sources, sentence by sentence
Archana is the research companion inside Citation Monster. She has read the paper and brings page numbers. When the sources don’t answer the question, she says so.
“Archana reads, checks, files. You decide.”
- Every sentence of an answer is checked: supported, partly supported, or not supported by the passage it cites.
- A page chip opens the paper at the exact words she quoted.
- When the paper doesn’t say, the answer is “Not enough evidence in these sources.”, with what she checked and where to look next.
- Anything she adds from general knowledge is labelled as such, apart from what the paper says.
- One chat for the paper, your library or the web; pick Fast, Thorough or Auto from the composer.
BERT: Pre-training of Deep Bidirectional Transformers
What batch size did they pre-train with?
Archana
256 sequences of up to 512 tokens, for 1,000,000 steps: “We train with batch size of 256 sequences (256 sequences * 512 tokens = 128,000 tokens/batch) for 1,000,000 steps” p. 13Supported
1 of 1 sentence supported · Quote found on its page
How much energy did pre-training use?
Archana
Not enough evidence in these sources.
I checked all 16 pages of this paper. It gives the hardware and the time, not the power draw:
- “Training of BERTLARGE was performed on 16 Cloud TPUs (64 TPU chips total).” p. 13
- “Each pretraining took 4 days to complete.” p. 13
Where to look next
The evidence layer
What the paper reports, each with its quote
While a paper is processed, its facts and claims are pulled out once, each with the words it came from and where they sit. Every run and dataset that includes the paper reuses them.
- Built as part of processing a new paper; nothing to click.
- The model proposes; a deterministic check decides. Each quote must be in the paper’s text, and a fact’s value must be in its quote. A quote found nowhere is marked failed and given no page.
- When two passages disagree, the conflict stays open until a person picks a value, with a reason.
- Facts, claims and conflicts by section, as charts you can filter.
Attention Is All You Need · Evidence
WMT 2014 translation
BLEU, EN→FR, big model
41.8 | 41.0
The passages disagree. Both stay on record until someone decides; the value you don’t pick is kept in the history.
- Abstract · p. 1Quote check passed
…a new single-model state-of-the-art BLEU score of 41.8…
- Results · p. 8Quote check passed
…our big model achieves a BLEU score of 41.0…
Grow this paper
From one paper to the work around it, on a plan you approve
Grow follows citations in both directions and finds similar work, then hands you a screening queue. Nothing runs until you approve a plan with a hard spend cap, and every decision is kept with its reason.
- A versioned protocol: question, inclusion and exclusion criteria, sources and a cap. A run starts only once a person has reviewed the criteria.
- The cap pauses a run before it would be exceeded, with the remaining work recorded; resume repeats nothing.
- Archana suggests; people decide. Screen by keyboard, judge each criterion against quotes from the abstract, and give a reason for every exclusion.
- Optional blinded dual screening, with a conflicts queue and agreement (Cohen’s kappa).
- A PRISMA 2020 flow (SVG and CSV), a discovery curve, and a methods appendix with every search, cost and decision.
- Re-run weekly or monthly; only what is new gets screened.
Grow run · protocol v1
What later work builds on, tests or challenges “Highly accurate protein structure prediction with AlphaFold”, and what did it find?
Scholarly databases · criteria reviewed before the run
Records · square-root scale
- Found457
- After removing duplicates382
- Ranked by Archana200
- For you to screen32
- Included9
Spend
$0.04
of a $0.50 cap
Screening · 1 of 32
Predicting multiple conformations via sequence clustering and AlphaFold2
Wayment-Steele, Ojoawo, Otten, et al.
Why it may be relevant
It tests and extends AlphaFold2 to predict alternative protein conformations, with experimental checks.
“We used nuclear magnetic resonance spectroscopy to confirm an AF-Cluster prediction” in the abstract
- I1 Reports original results: met
- E1 Survey or review: not met
I IncludeE ExcludeU Uncertain1–4 ReasonLeft arrow Undo
Datasets, synthesis and API
A dataset you can trust, cell by cell
Turn what your papers report into a typed table. Every value opens the quote and page it came from, and a gap stays a gap.
- Typed columns: text, integers, yes or no, numbers with their units. Start from a table in the paper, a proposal built from its evidence, or your own columns.
- “Not reported” is never zero, and “not applicable” is its own state. Conflicting values stay open until someone decides.
- Snapshots freeze a dataset; they can’t be changed afterwards, and export to CSV or JSON with a data dictionary and manifest.
- A grounded synthesis of each snapshot: every finding cites its rows and quotes, null and contradictory results are kept, and interpretation is labelled.
- Notebooks and agents read snapshots through a read-only REST API or MCP, with workspace-scoped tokens you can revoke.
Read-only API and MCP
# a token with dataset:read
GET /api/v1/datasets
GET /api/v1/snapshots/{id}/rows
GET /api/v1/snapshots/{id}/provenanceWMT 2014 results · snapshot v1
| Resulttext | Yearinteger | EN→DEBLEU | EN→FRBLEU |
|---|---|---|---|
| ByteNet | 2017 | 23.75 | Not reported |
| GNMT + RL | 2016 | 24.6 | 39.92 |
| ConvS2S | 2017 | 25.16 | 40.46 |
| Transformer (base) | 2017 | 27.3 | 38.1 |
| Transformer (big) | 2017 | 28.4 | 41.8 | 41.0, conflict |
Cell · Transformer (big) · EN→FR
…our big model achieves a BLEU score of 41.0…
Results · p. 8 · disagrees with the abstract (41.8) · open until someone decides
Trust and privacy
Unpublished research stays yours
Citation Monster holds work that isn’t public yet. The rules are short, and the privacy policy has the detail.
Your work never trains AI models
Not your uploads, highlights, notes, chats, questions, research runs or knowledge bases. Not ours, and not our AI providers’.
Private to your workspace
What you upload and create stays in your workspace. Only its members can see it, and the database itself enforces that.
What leaves, and why
To do the work you ask for, the parts of a paper a task needs go to an AI provider (Anthropic, OpenAI) or our OCR provider (Mistral), under terms that forbid training on them.
Costs before they happen
Model work counts against your plan’s allowance. Grow runs stop at a cap you set, and the run view shows what was spent.
Also in every workspace
The rest of the reading room
Graph
Knowledge graph
How the references connect: which cite each other, which are cited together, which share sources, and the topics they cover.
Scholars
Scholar pages
Author names open a page with their identifiers, affiliations and works.
Knowledge
Knowledge bases
Collect papers and notes into a base Archana can read, and connect it to other tools over MCP.
Interface
Light, dark or classic
The interface follows your system’s light or dark setting. The classic interface is one switch away in your account menu.
Pricing
Start free; pay when the work grows
Reading, notes, citation graphs, checks and exports are free forever. Plans add extracted papers, chat, research runs and knowledge bases. Every paid plan has an academic price for a verified .edu email.
Free
$0 forever
1 extracted paper a month, plus everything free forever.
Researcher
$20 a month
For one researcher reading steadily.
Researcher Pro
$50 a month
For heavy reading and Thorough-mode work.
Lab
$30 a month per seat
For a group: per seat (3 to 50), limits pooled.
FAQ
Questions
How it works, what it checks and what happens to your papers.
What does Citation Monster do?
You upload a paper as a PDF and read it as a page, with every citation marked where it sits: verified, differs, retracted or not found. Ask Archana about it and each sentence of her answer shows its source. The paper’s facts and claims are pulled out with their quotes, Grow finds and screens related work on a plan you approve, and datasets turn what your papers report into a table where every cell opens its quote.
How are references checked?
Each reference is checked against the published scholarly record, and one that isn’t found there is checked on the web. Verified means the work exists as cited; differs means the title, authors, year or venue don’t match (the corrected details are shown); retracted shows the notice’s date and source, and retractions are checked again weekly; not found means no matching work turned up, which is often a sign of a fabricated citation. Treat the results as a first pass and open the linked source before you rely on it.
What does “Audit this paper” add?
It reads every citing sentence and says how the paper uses the work there (supports, contrasts, mentions, or uses its method or data) and whether the work says what the sentence attributes to it: supported, partly, not found in sources checked, or contradicted. It reads the cited paper itself when it is in your workspace. Every supported answer quotes the passage; without one, the answer is “not found”, never “false”.
How do I know an answer from Archana is right?
Each sentence is checked against the passage it cites and marked supported, partly supported or not supported, and a page chip opens the exact words. Anything from general knowledge is labelled. When the sources don’t answer, she says “Not enough evidence in these sources.” and what she checked. Answers can still be wrong, so check the quoted pages.
What does Grow cost, and who decides what goes in?
You do. Grow drafts a plan (question, criteria, sources) and a hard spend cap; nothing runs until you approve it, and a run pauses before it would pass the cap. Archana suggests; people include and exclude, with reasons, and can screen in pairs, blinded. Model work counts against your plan’s AI allowance.
Can I use my datasets in other tools?
Yes. Export a dataset as CSV or JSON with a data dictionary and a manifest, or read its frozen snapshots from a notebook or an agent through the read-only REST API or MCP, with an access token scoped to the datasets you choose.
Which AI models does it use?
Claude, from Anthropic, does most of the reading and checking; GPT models from OpenAI are available in chat; Mistral’s text recognition reads page layouts and scans. Archana picks the model for each answer in Auto, or you can choose Fast or Thorough.
Is my data safe?
Papers you upload are stored in your workspace, and only its members can open them. To do the work you ask for, the parts of a paper a task needs are sent to the AI providers that run those steps. Neither we nor they use your papers, notes, chats or questions to train AI models. Details are in the privacy policy.
Can I keep the old interface?
Yes. The new interface is the default; switch to the classic one from your account menu or Settings, and back again whenever you like.
My file won't upload. What should I check?
Uploads must be PDF files up to 50 MB. A scanned PDF without a text layer is read from the page images, so very faint or handwritten pages may come out incomplete. If a file fails, the upload shows the reason and lets you try again.
Bring the next paper you have to read
Upload a PDF, see which citations hold up, and keep what you find.