PairProof

PairProof — AI writes the code. You score how the engineer directs it. Traditional interviews ban the tools people use at work. PairProof flips that: a live, AI-native technical interview where the candidate is scored on approach selection, intervention, verification, and token efficiency — not typing speed. One Rust CLI, two roles. The expert pastes a job description (Meta extracts topics/level/language), hosts a disposable Docker sandbox, pulls LeetCode-style questions, and shares a rotating connection string. The candidate joins with Mutagen + SSH, picks among Meta-generated approaches, steers the model through continue/redirect/change/test/finish, and works in a synced local workspace (q1/, q2/, …). Scoring is explainable by design: a fixed live step table (starts at 6.0, clamped [3, 10]) plus a capped Meta grade at close (±0.8). Chat and failing tests don’t inflate the live score — the AI wrote the code. The session ends with sessions/<id>/report.md for decision support, not automated hiring. Built on the Meta Model API (muse-wonder-1.0 / muse-spark-1.3) for extraction, approaches, coding, coach chat, and grading — all expert-side only. API keys never enter the sandbox or connection string.