The demo's judge is still a placeholder, so this week we started replacing it, measurement first.
New: a judge client for an open model we serve on our own GPU. No contributor photo goes to a commercial AI API. Server errors and malformed replies are held, never paid.
We also built a labelled set of 232 cases from our own photos: good, generic and mismatched descriptions, plus prompt-injection attempts. The placeholder judge agrees with the labels 31.5% of the time, because it pays everything. That is the baseline a real model has to beat before it decides any payment.
Next: run the model and publish the number.
Give feedback to the team
Your message is private to you and the Survey2Earn team.
Create an account to send feedbackAfter signup, you can write a private message with this update attached.