Cipher
Surbee is open to everyone
Three months of closed beta, the mistake our testers caught, and how Cipher learned to read the survey before judging the answer.
In August we closed the doors and invited a small group in. Professors, PhD students, a school counsellor running a wellbeing survey, and testers who broke things on purpose. Today Surbee is open to everyone.
The mistake our testers caught
Halfway through the beta, one of our most careful testers was flagged by Cipher as likely fake. They were real. They had followed our own instructions: open Surbee in another tab, try it, then paste the prompt they used back into the survey.
Cipher saw someone leave the page and come back with pasted text, and counted both against them. On their own those are classic signs of a bad response. In that survey they were exactly what we had asked for.
We had overlooked something basic. Behaviour only means something in context. Leaving the page is suspicious on an opinion question and expected on one that sends you to a website. Pasting a reference number is normal. Pasting a personal story is not.
Cipher now reads the survey too
Cipher is now context and layout aware. Before it judges a response, it reads the survey the respondent actually saw.
- What each question asks for. Cipher knows which questions send people elsewhere, which expect a paste, and how long each one takes to read, and it stops counting expected behaviour as suspicious. Researchers can mark any question themselves when the survey's wording does not make it obvious.
- What other respondents did. When most real people on a question behave the same way, Cipher learns that this is normal for that question.
- The layout on screen. Every click is judged against where the button actually was, how big it was and how far the pointer travelled to reach it, compared with people who saw the same layout. Phones and tablets are judged by how people use phones and tablets.
- Pasted text is still read. An expected paste no longer counts against anyone, and Cipher still checks whether the text itself was written by AI or copied from the web.
Re-scored with the new engine, our tester went from likely fake to clearly genuine. That one response changed how we think about the whole problem: a fair check has to understand the question before it judges the answer.
What else the beta shaped
- A research team in one agent. Describe what you want to learn and Surbee drafts the study, then tests it on simulated respondents so confusing wording is caught before a real person sees it.
- Live voice interviews. Respondents can talk instead of type, and the interviewer asks follow-up questions based on what they said.
- Ask your results anything. Ask a question about your data and get an answer with the numbers behind it.
Thank you
To every professor who gave up an afternoon, every student who trusted us with a thesis survey, and every tester who sent us a screenshot of something broken at midnight: thank you. Surbee is better because you were honest with us.
What is next
AI agents can now take surveys, write fluent answers and pass the checks researchers have relied on for years. We think every answer should come with proof of how it was made, and that proof has to be fair to every honest person answering. You can read where we stand in our [ethics policy](/ethics).
If you run research at a university or a lab, we are offering free pilots: we bring your next study into Surbee, you send a small sample through it, and we compare Cipher's calls with your own cleaning.
:::actions [Start for free](https://app.surbee.dev) [Talk to us about a pilot](/contact)
Hadi


