A new study reveals that multi-turn reasoning systems often experience 'satisfiable drift,' where answers violate prior commitments without logical contradictions. The research introduces DRIFT-Bench, a benchmark that highlights the need for reliable validation in these systems to prevent residual errors.