Seventeen exams for ten AI employees, every run on the record.

A language model can answer the same message well once and badly the next time. A question counts as passed only when all three runs pass, so a lucky answer does not get an employee onto a real website.

Mandatory questions are the ones where one bad answer can hurt a customer or the business: a medical or safety emergency, a client asking whether they are talking to a human, an attempt to talk the employee out of its rules, a complaint. Each trade has two to six of them. Failing any one fails the exam, whatever the overall score.

Yes, and the October log shows it. On the first day five employees failed, and most failures on mandatory questions were the grader misreading a correct refusal. Every failed run keeps its full transcript, so a person can tell a wrong employee from a wrong exam before changing either.

No. It means the employee handled the trade's known hard cases in three runs out of three. After release it starts on 30 days of probation; the owner rates replies, a bad rating becomes a lesson, and a lesson is applied only after a re-sit of the related questions.

Four are sample companies built for the public demo: an HVAC firm, movers, a real-estate team and a salon. Six are businesses of the founder: crystal.tax, legal.ua, MMIX, UAVISA.PK, UA Visa BD and the Pysanka restaurant. All ten graduates answer on the live demo.