A chat on the file
25%
Would delist because of the hygiene score. 25 out of 100. That’s the question in the file.
Fieldlock
Ask your surveys in English. Where a question was not asked, you get no number.
Your client wants one more number before the call. Type the
question the way they said it; back comes the count, the wording
it was counted from, how many people, and a receipt. Where that
question was not asked on the wave, the card stays blank, names
the look-alike it refused, and shows the figure a chat would have
given you. No language model reads the file or writes a figure;
every receipt says model=none. The file
stays with you, in Frankfurt, and your client reads the cards on
a page of their own.
In-house teams and agencies, on live studies
- Octapharma
MarketCulture
This is the real engine, counting on our server in Frankfurt on a made-up study. Harborline is a food packing survey with two seasons. Look at the Next season card first. That 25% is the figure a chat on this file would have counted. Fieldlock shows it as not counted and leaves it out. Then ask it anything.
NextFluent Fieldlock
Public sample
Sign in
Harborline packing
/
Ask
Counting 100 rows on the desk.
A chat would tell you 25%.
Last season you asked retailers why they delisted. Twenty out of a hundred said the hygiene score. Next season that question was not asked. The file does hold a look-alike: would they delist if the hygiene score were poor. Twenty-five out of a hundred said yes. A chat on the file counts that one and tells you 25%.
Fieldlock
No figure
We put that question’s name on the page, show its 25% as not counted, and leave it out.
Last season, the question you asked is in the file. 20 out of 100 said yes. Unweighted. The wording sits under the number, and the receipt under that. We ran the question past open models on 9 November 2026: the largest gave 25% in four runs of four. Read the test.
The assistants now built into survey platforms compute with a statistical engine, so their arithmetic is exact. So is 25 out of 100. Exact arithmetic does not tell you it was the question you asked. What to ask of any tool.
Why research teams pick Fieldlock.
It refuses, and shows you the figure it refused.
Where the question you asked is in the wave, you get the percentage. Where it is missing, the card stays blank, names the look-alike it would have been given by mistake, shows that figure as not counted, and prints the wording that makes it a different question. You see the 25% a chat would have handed you, and the question it really answers.
No language model near a number.
Every figure is arithmetic on the rows of your file. Your
English is bound to a question by scoring it against the
wording and the answer labels in the file. No model reads the
file or writes a figure, and the receipt under each figure says
model=none, there for your client to read as
well. One thing a model may do, once the owner of your desk
switches it on: read the questionnaire you fielded, so the card
carries the question as the respondent saw it. It reads words
and never sees a row.
A receipt your client can check themselves.
Under each percentage sits the wording it was counted from, how many people answered, a hash, and Fieldlock’s signature over the hash. Your client reads the card on the page you give them, and anyone holding the receipt can paste it at app.nextfluent.com/verify, with no login, and watch it check. Count it again in a year and it says whether it still stands.
Do not take our word for it.
An open test, for any tool.
39 questions on three made-up studies, built around the look-alikes real files carry, with an answer key you can recount in a spreadsheet. Put it to whatever answers your survey questions today. Fieldlock sits it on every build, and a build that gives one wrong number does not ship. Take the trap pack.
A public log of every receipt.
Each receipt that leaves a desk goes into a log anyone can read: its hash and the time, nothing else. The log is chained and signed. Keep a copy, and you can see if a receipt is later withdrawn, altered or back-dated, by us or by anyone. Read the log.
A number that answers for itself later.
A receipt names the engine that wrote it. Any receipt on a desk can be counted again today, and it says whether it still stands and, where it does not, whether the file, the codebook or the engine moved. Before a card reaches your client, the server holds its figure to its receipt.
The desk, as it is.
Your studies down the side, your questions in a thread, one card per wave, and under each card the steps that produced it. The screens below are the real desk on the Harborline sample and on a second made-up tracker with two waves.
Your surveys, on one desk.
The sample above is one made-up study. The desk is where your real studies live, and where the people who work them sign in.
Load the studies
Drop the SPSS file. It carries the question wording and the answer labels, so there is nothing to type in. A CSV works too. Add the study you are working on, or every study you run.
Ask one question across all of them and each study answers on its own. We never blend two studies into one percentage.
Ask in English
Type the question the way your client says it on the phone. Back comes the figure, how many people, and the wording we counted. When a question was put to some people only, the card says who, worked out from the file itself: “Asked only of the 484 who ticked the brand on the awareness question.” A weight in the file is applied and named. A 1–5 question comes back as a named box or a mean. An amount comes back as a mean, or as the share over €50, with the refused codes left out. A ranking question comes back as a table of first choice, top three and mean rank. Ask what drives a score and the card carries the drivers table.
Add “by plant” to cut it, with a letter on each column and the 95% test between them. Add “among those who changed supplier” to narrow it. Ask whether something changed since the wave before and the change comes back as a card of its own: the points moved, and whether that is more than chance on these bases. With three or more waves it draws the line.
When a new wave lands, the desk writes the debrief itself: every question asked on both waves compared, the moves beyond chance listed largest first, each with a receipt. Under each move is what the file can count about why: which groups carried it, whether the base changed, which rated items moved with it. Where the file holds nothing, the card says so and the question goes on the list for the next questionnaire.
Give your client the page
One page per client, each with its own link, its own studies and its own cards. They read what you sent, and they can ask for themselves. You see every question they typed before the next call.
List their work email and they sign in once to find every page they have been given. In-house, the page goes to the brand team down the corridor, and the file stays with research. See a client page.
What a chat would get wrong on your file.
Your own file goes on the desk, not on this page. Book a 30-minute call and we open a desk with your file on it: an SPSS .sav or a CSV from any survey tool. It lists what a chat on that file would get wrong: the look-alike questions, what was dropped between waves, the refused codes in an amount, the weight. Then you ask it. The file stays in Frankfurt, behind your sign-in. Do not email the file.
The desk keeps your studies, remembers your questions, lets you pin wording, and gives your client a page. Open a desk for your file. See pricing.
Who sits at the desk.
In an agency, the researcher and the account lead sit at the same desk. In-house, the insight manager and the marketer who keeps asking. Each has their own password and authenticator, and the log says which of them sent the card. A colleague is €600 a year, agreed in writing before they sign in.
The researcher
You run the study and you answer for the numbers. The desk gives you a figure with its wording and base in the time it takes to type the question, and refuses when the honest answer is no figure. Pins keep your decisions.
The account lead
You are on the phone when the client asks. Ask the desk the way they said it, copy the card into the deck, and send it to their page. Every question they typed themselves is listed before the next call.
The client
You get one page from the team that ran the study, your agency or your own research department. On it, the cards they sent and a box to ask your own question. Every number carries the question it came from and a receipt you can check. The file stays with them.
Three things people ask first.
Which files work?
An SPSS .sav, which brings its own wording and labels. Or a CSV, with a two-column labels file if it has none. A weight in the file is applied and named. More on files.
Where does my file go?
Onto your desk, on a server in Frankfurt, under your login. Nothing from it is sent to a language model. Your client sees the cards you send and the answers to their own questions; the file itself stays with you. Security and data.
What does the receipt prove?
That this figure came from this study, this wave, this question, this field and these rows. Run it again in a year and the hash matches. How we count.
Start with one study.
The desk is €2,400 a year for one seat. With a page for one client it is €3,600. A further seat is €600 a year and a further client is €1,200, both agreed in writing. We set up your first study with you, on a call, in an afternoon.