How privacy works
Three rules, and one diagram of what actually happens.
Your paper is never uploaded.
Not for matching, not for the format check. The file stays on your device the whole time.
Nothing from a paper is stored anywhere.
Not on a server, not in an account — even if you sign in later, paper content is never part of what's saved.
Anything that would need to leave your device is opt-in.
Matching, the format check, and the journal rules check never do. One feature is the exception: getting a paper reviewed uses a large language model, which can't run in a browser — it asks first, in plain language, exactly what it's about to send, before sending anything.
What actually happens
The embedding model and the journal index are public files — the same ones for everyone, downloaded once and cached by your browser. Your paper never appears in either direction of those downloads. There is no MargaLink server in this loop at all: matching a paper to a journal is a page in your browser talking to files, not to us.
Why this is checkable, not just claimed
Open your browser's network tab (or use the panel that appears on the upload page after you run a match) and read the requests yourself: every one is a plain GET for a public file, none carries a request body. The browser-side code that does the extracting, embedding, and ranking is what actually ships to your browser — there's nothing hidden behind a server to take on faith.
What an embedding reveals
Matching works by turning your paper's title and abstract into a list of a few hundred numbers (an embedding) and comparing it against the same kind of list for each journal. Be clear about what that number list is: it carries the general topic of your paper, not the words. It can't be turned back into your original text in any practical way — but it's still derived from your paper, so we don't send it anywhere either. The whole comparison happens locally, against the journal index already in your browser.
The one exception: getting a paper reviewed
Everything above — matching, the format check, the journal rules check — runs entirely on your device. Getting a paper reviewed is different: it sends your paper's text to Anthropic's Claude API, because that kind of review needs a large language model, and no model capable of it runs in a browser today. This is the only feature in MargaLink that works this way, and it only runs if you explicitly ask for it:
- Nothing is sent until you confirm a plain-language notice naming exactly what's about to happen — there is no default-on path.
- Author names and email addresses are stripped from the text first, on a best-effort basis, before anything leaves your device.
- Anthropic's API terms don't use this data to train models. MargaLink doesn't store what you send, before or after the review.
- Like every other request in MargaLink, this one shows up in the network-request log on the matching page when it happens — it isn't hidden from the same transparency check the rest of the site relies on.
Accounts and payment
If accounts exist later, they'll need only an email address. Payment details stay with the payment provider — MargaLink never sees or stores card numbers.