NextFluent

Security and data

Your file stays on your desk. Your client gets the page.

A response file is the most sensitive thing a research team holds. This page says, plainly, where it goes, who can see it, and what leaves your browser. The legal wording is on the trust note and the data processing agreement.

Where things live.

The sample

The Harborline sample on the homepage is a made-up study. When you ask it a question, the question goes over HTTPS to our server in Frankfurt, is counted there, and the card comes back. The question is not stored. There is no upload on the website: your own file goes on the desk, behind your sign-in.

The desk

A study you add on the desk goes over HTTPS to the EU. The database, sign-in and file storage are on Supabase in Frankfurt. The application is hosted on Vercel, with its server code in Vercel’s Frankfurt region. Each organisation’s rows are isolated from every other’s.

The client page

Your client’s page carries figures, wording, bases and receipts, and the questions they type. It never carries the response file or the codebook, and the file is not in the link. Each client has a page of their own; nothing from one page shows on another. A page can be opened to listed addresses only: the link alone then opens nothing, each reader signs in with a link sent to their own inbox, and the database itself turns away anyone who is not on the list. New pages start that way.

No language model on the number path.

Every figure Fieldlock shows is arithmetic on the rows of your file. Binding your English question to a field is done by scoring against the question wording, the answer labels and the notes the desk took when the study was added. No language model is asked for a number, a field or a sentence.

Each receipt records this as model=none. Your response file is never sent to ChatGPT or to any other language model provider. The companies that process your study data are Supabase and Vercel. Sign-in emails leave through Resend, which sees an address and nothing from a study. We will tell you in writing before adding one.

One thing a model may read, and only when the owner of your desk has switched the reader on under Account: the questionnaire you give a study. Its text and the codebook’s labels and answer codes go to Mistral AI, a French company whose API is served from the EU, for that one reading, and the model proposes the wording each question had. Nothing from the response file goes with it. A gate drops anything in the answer that reads as a figure, and a person on your desk accepts the wording before the codebook changes. Switching the reader on is dated and logged in the owner’s name; switching it off ends it.

The desk does learn, and only from wording. When you pin a question to a field, or send a card to a client, the desk keeps the generic English of the question (how, many, have, ever, bought, would, among, last, top, box; a closed list) and counts it against the kind of field you bound it to. A phrasing that two or more desks land on the same way becomes a cue the engine reads when its own lists are silent. No row, no label, no name of a product, person or client is part of it, and a desk can keep its wording to itself from the account page. Bilingual Belgian files are read in Dutch and French as well as English, so an English question binds to “Hoe tevreden bent u” with the same receipt, and on such a file the Ask screen says you can ask in Dutch or French as well.

In transit
TLS on every connection, from your browser to the desk and from the desk to the database.
At rest
Encrypted at rest with the host, in Frankfurt.
Sign-in
Invite only. There is no public registration. You write to us, we open a login, you upload in the desk. Never by email. Every desk login carries a second factor: a password and the six digits from an authenticator app.
Isolation
Row-level isolation per organisation in the database. A study on one desk is not visible on another. The database itself refuses a session that has not given the second factor.

How the desk is locked.

Two things to sign in

A password opens nothing on its own. Each member of a desk sets up an authenticator app on first sign-in and gives its code every time after. The rule is enforced in the application and again in the database, so a stolen password reaches no rows. Lost your phone? Write to us from the address on the account and we take the authenticator off it after a call.

A colleague joins the desk when its owner types their work email on the Account page. The desk mails them a link; they choose a password, set up their authenticator, and sign in on their own from then on. The owner takes a colleague off the desk from the same page. The login stays, with nothing to open. Nobody else can add a seat.

What the browser may run

Both sites send a Content-Security-Policy. On the desk, a script runs only when it carries the one-time nonce the server minted for that page; on this site, only scripts from this site run. Neither page can be framed by another site. Connections stay over HTTPS for two years by header, and the sign-in cookie is out of reach of page scripts.

Kept up to date

The code behind the desk is checked every week for published vulnerabilities in what it depends on, and fixes land as they appear. The build fails on a known problem in a package that ships to production.

The log

Every action on a desk is written to a log in the name of the member who did it: a sign-in, a question asked, a CSV taken, a card sent to a client page, a client link remade, a client page made or removed, a colleague added or taken off, a study or wave deleted. Members read their own organisation’s log on the desk. Nobody edits a line and nobody deletes one, and the lines outlive the study they name.

The client page under load

A client link answers a bounded number of questions a minute and serves its page a bounded number of times a minute. Past that it asks the visitor to wait. The desk has the same questions, from the same link, listed for you to read.

Who sees what.

You, on the desk Your client, on their page Anyone with a receipt
The response file Yes Never Never
The codebook Yes Never A fingerprint only
Figures, wording and bases Yes The cards you sent, and answers to questions on studies you opened What is in the receipt
Questions your client types Every one, listed on the client page desk Their own No
Receipts Every one On every card Can re-hash it at /verify without a login

How long we keep it, and how it ends.

While the desk is paid for

Study data is kept for the paid year of the desk: the fields, pairs, pins, waves, receipts and the response files behind them. You delete a wave or a whole study yourself, from the Waves tab of the desk, by typing its name. It goes at once, with its rows, receipts, pins and client cards. The log keeps one line saying who deleted it and when.

When the desk ends

We keep study data until you ask us to delete it, and then we delete it within 30 days, except records the law requires us to keep. Ask for a copy first and we return it before deleting. Write to trust@nextfluent.com.

Where the company stands.

Security questions and reports: trust@nextfluent.com.

Bring the question, not the file.

Write to us from your work email and we open a login. You upload in the desk, on a call with us, in an afternoon. Do not attach a file to the email.