Ai Led Interviews And Scoring

Mercor markets ~20-minute AI-led interviews. A candidate's interview cuts off at minute 12 mid-answer because the conversational agent decided it had enough signal.

Interview length is a candidate-trust surface — early termination must follow a documented criterion (signal saturation, candidate disengagement, technical fault) surfaced to the candidate with a re-take option when caused by Mercor. Do not silently truncate a candidate's response. [REQUIRES-VERIFI…

Two candidates give similar answers to the same interview question on different days. One scores 4/5; the other scores 2/5. The score difference comes from a drifted rubric anchor in the AI grader.

Grader rubric anchors must be versioned and frozen per cohort; any rubric change must apply only to interviews from that change forward, with a documented version ID stored alongside the score. Periodic calibration runs verify that anchor interpretation is stable across time.

An AI grader is trained on a US-English-dominant calibration set. A non-native-English candidate gives a technically correct answer with grammatical errors and receives a lower score than a less-correct US-English candidate.

Grader must score on substance, not surface fluency. Run periodic adverse-impact audits across English-fluency cohorts and publish the selection-rate ratio to compliance. If a fluency penalty is found, retrain or post-correct rather than ship the biased score to the customer lab.