Causal Audit Reveals Synthetic Distractor Biases in RLVR Data

Auditing the GooseReason-0.7M dataset reveals that RLVR pipelines using synthetic distractors can introduce unintended shortcuts. Specifically, code distractors often manifest as simple operator mutations of the gold answer rather than independent alternatives, potentially training models to recognize artifacts rather than reasoning.