What is versim?
versim turns real online communities into queryable personas. You ask a panel of typed personas a question and get back anonymized answers in the community's voice, with a score for how realistic the answers are.
faq · plain answers
Plain answers to common questions.
What versim is, how it differs from a generic LLM, where the data comes from, what the fidelity score means, how to opt out, and how to join the waitlist.
Jump to a section
the product
What versim is and how it works
How is versim different from using a generic LLM as personas?
A generic LLM pretending to be a senior engineer is a guess from a job title. versim's personas are built from real people's actual messages: their tools, their voice, their opinions. We test the personas against real messages they never saw and publish the score. Generic LLMs can't be tested this way because there's no specific real person to compare against.
What does the fidelity number mean?
A score from 0 to 20. An anonymous judge compares each persona to real messages it never saw, scoring four things: speaking style, specific knowledge, voice signature, and factual accuracy. Higher means more realistic. Our current grounded score is 17.80 out of 20 versus 10.60 for the same model without grounding. See how we measure fidelity for the full method.
data and privacy
What we do with the source data
Where does versim get its data?
From community archives we have permission to use. Either open community archives that the community has chosen to make publicly accessible, or direct partnerships with community owners. We don't scrape private channels and we don't buy or trade scraped data.
Does versim sell my community's data?
No. The source archive stays with us. We don't sell, share, license, or publish the underlying data. The product is the anonymized panel output, not the data behind it. See how we treat the data for the full posture.
Can my community opt out?
Yes, at any time. Email hello@versim.ai from an address connected to the community. We delete the source archive, all dossiers built from it, any cached panel output, and any fidelity scores derived from it. Promptly: days, not months. We confirm completion in writing.
Will real people in the source community see my questions?
No. Real people never see your questions. You ask the synthetic panel, which is built from the real archive but runs entirely on our infrastructure. The real people are not contacted and don't know a query happened.
Can someone identify a real person from a panel response?
No real names, handles, or contact information ever reach you. Exact source text is stripped at fifteen words or more. Per-account question limits prevent reconstruction by repeated querying. The anonymiser is enforced by code and tested on every release.
getting started
Joining and pricing
Is versim available right now?
We're in pre-launch. The waitlist is open. The hosted product is in development. The methodology and the scores published on this site are real and repeatable.
How much will versim cost?
Pricing isn't set yet. We're talking to early customers to figure out what makes sense. If you're interested, get on the waitlist and we'll bring you into the pricing conversation.
What kind of communities does versim work best for?
Online communities that talk publicly: open Telegram channels, Discord servers, subreddits, mailing lists, forums. The richer the archive, the more voice texture the personas will have. Smaller or quieter communities work too, but the personas will have less to draw from.
hard questions
Two questions worth asking us
What if the judge model is biased?
The same model family judges both the grounded and ungrounded sides, so any model bias should affect both equally. The seven-point gap (17.80 versus 10.60 out of 20) survives that. Cross-model judging, using a different model family as the judge, is on the roadmap. We expect the gap to hold but won't claim it until we have the data.
Why should I believe your fidelity number?
Because we publish it and the test is repeatable. We share the full evaluation (the questions, the raw judge scores, the prompts, and the runner) with serious customers. We'd rather you run the test yourself than take our word.
Related
- How we measure fidelity — the fair test that scores a persona against real messages it never saw.
- How we build a persona — the end-to-end pipeline from real archive to panel answer.
- How we treat the data — the full privacy posture and community rights.
- Back to versim — the landing page, with the sample panel run and the demo.
Get in touch
Question not answered here? Write to hello@versim.ai.