A chat on the file
It writes the brief well, and it puts a percentage in it. That percentage may come from the question next to the one you meant, and you cannot tell from the brief. Your file is now in the chat.
Compare
Ask a chat the Harborline question about next season. That question was never asked in the next-season wave. The chat answers anyway, from the nearest column it can find, and its arithmetic is right: 25 out of 100 is 25%. It is the answer to a different question. Fieldlock leaves the card blank, shows you the 25% it refused to count, and tells you which question that 25% belongs to. Then it signs what it did, so anyone can check.
On 9 November 2026 we handed open models the Next season wave of the Harborline sample, variable labels and 100 rows, and asked the client’s question: “What percentage of retailers delisted last season because of the hygiene score?” That question was asked last season, 20 of 100. It was not asked next season; the file holds “would delist because of the hygiene score”, 25 of 100. Each model was asked in prose on the file and as code it wrote, which we then ran on the real rows, up to two runs of each. We used Mistral’s models because anyone with a key can run the same script and read the same answers.
| Model | What came back |
|---|---|
| Mistral Large 4 | 25% in all four runs. In prose it noted that the file has no last-season item and gave 25% anyway, by plant. As code it printed 25.0% and said nothing. Given both waves and asked why, it found the 20 of 100 on last season. |
| Ministral 14B | 61.5% in prose, under a worked method. As code, 50% twice, counting the retailers whose main reason was hygiene. |
| Ministral 8B | 0% and 30% in one answer, 12% in the next. As code, once an error and once 0.00%. |
| Fieldlock | Last season 20% of 100, unweighted, with the wording and the receipt. Next season no figure, with the 25% named as not counted. Each card carries its receipt, checkable at app.nextfluent.com/verify. |
Asked why retailers delisted, every model wrote an explanation with figures in it, from a wave on which the question was not asked. The large model built its case on a coincidence in the synthetic rows, that every lot ticking the hygiene score also expects a quality hold; the smaller ones told a different story on each run. Read every answer as it came back: the test note, the script, and the two files the models were given, last season and next season. Both are the synthetic Harborline sample; no client data is in them.
It writes the brief well, and it puts a percentage in it. That percentage may come from the question next to the one you meant, and you cannot tell from the brief. Your file is now in the chat.
The platforms that build your tables now let you ask in English too. The careful ones say the figure is computed by their statistical engine and not written by a language model, and that the assistant says so when the data cannot answer. That is real progress, and it settles the arithmetic. It leaves the step before the arithmetic: which question in the file your sentence was taken to mean. A figure can be computed exactly and still be the look-alike’s.
We have not run those assistants. They sit behind accounts we do not hold, and we will not print a result we did not get. So we published the test instead of a verdict: the files, the questions and the answers. Put it to the tool you use, and to us. The trap pack.
Last season is 20% of 100, unweighted, from retailers who delisted because of the hygiene score. Next season is blank, with the look-alike question and its 25% shown as not counted. Run it again a year later and the receipt matches. Your client can ask the same study from their own page. Keep the chat for the brief and your analysis platform for the slides. Fieldlock is where the number comes from.
Whatever answers your survey questions, four things can be asked of it. Each of these is public, and none of them needs a login.
39 questions on three made-up studies: 18 with a look-alike in the file, 6 on a wave that did not ask them, 7 the study never asked. Every answer in the key is a recount of rows you can open in a spreadsheet, and beside it is the figure a tool gives when it falls for the look-alike. Fieldlock sits it on every build; one wrong number and the build does not ship. Take the pack.
Under every Fieldlock figure is a receipt: the question it was counted from, the people, the counts, a hash, and Fieldlock’s signature over the hash. A receipt someone wrote themselves hashes correctly and still fails the signature. Every receipt that has left a desk is in a public log, chained and signed: anyone who keeps a copy of the log can see if a receipt is later withdrawn or back-dated. Check a receipt. Read the log.
A receipt names the engine that wrote it, and every engine is on a public list. On the desk, any receipt can be counted again today. It says whether the receipt still stands and, where it does not, whether the file, the codebook or the engine moved. A card sent to a client is held to its receipt by the server before it goes.
| In the meeting | A chat on the file | Fieldlock |
|---|---|---|
| Wrong question | Counts a nearby question that happens to be in the file. | Counts the question you asked. If it is not there, there is no percentage. |
| Wrong number | Can invent a percentage, or mix two answer lists together. | Counts that question only. Run it twice and the receipt matches. |
| Next season | Answers from whatever is in the new file. | Will not copy last season’s 20% onto a question that was not asked. |
| Several studies | One prompt becomes one blended percentage. | Each study answers on its own. Nothing is averaged together. |
| The weight | Ignores the weight column, or applies it without saying so. | Applies the file’s own weight and names it. The unweighted figure sits beside the weighted one. |
| A 1–5 question | Picks a reading for you and calls it “satisfaction”. | Names the box it counted: top 2, bottom 2, or the mean on 1–5. Don’t-know codes stay out. |
| “Average spend” | Averages the column, 999 for “refused” included. | A mean on the people who gave a number. The refused and don’t-know codes are named and left out. |
| “Up from last wave” | Says so if the number went up. | Tests the change on the two bases at the 95% level and says whether it is real or within what chance would give. With a weight, the test uses the effective bases. |
| “Why did it move?” | Writes a reason. It reads well and nothing in the file was counted to get it. | Cuts the move by every classification in the file and tests whether any group is apart from the rest; splits it into change inside the groups and a changed mix of them; checks whether the base changed; names the rated items that moved with it. Where the file holds nothing, it says so and puts the question on the list for the next questionnaire. |
| “Why only 484?” | Shows the base and leaves you to find the questionnaire. | Works the routing out from the blanks in the file and says who was asked: “Asked only of the 484 who ticked the brand on the awareness question.” |
| A look-alike question | Picks one and tells you the number. | Counts the one you asked. Then it counts the look-alike as well, and tells you if the figure depends on which one you meant. |
| “By region” | Writes a table, fills a group of three people with a percentage, and leaves the reading to you. | Cuts the count by that question. Each column gets a letter, and the card marks which columns it is higher than at the 95% level. A group under five people is held back and says so. |
| Your client | Gets a deck, and emails you when they want one more number. | Gets a page where they ask for themselves. You see every question they asked. The file stays with you. |
| Your colleagues | Each one uploads the file to a chat of their own. | The file is loaded once. Each colleague signs in to the same desk with their own password and authenticator, and the log says who sent which card. |
| “Says who?” | A sentence in a chat. Nothing to check it against, and tomorrow’s answer may differ. | A signed receipt at its own link, in a public log. Counted again in a year, it says whether it still stands. |
| The file | The chat keeps the upload. | The percentage is not written by a language model. Invite only. Stored in the EU. |
€2,400 a year for the desk. €3,600 with a page for one client. Or ask the Harborline sample on the homepage first and read what a chat would get wrong, free, without a login.