$$R=R_{\text{abstain}}+R_{\text{company}}+R_{\text{time}}$$

• $R_{\text{abstain}}$ measures abstention accuracy, model gets full credit if it correctly decides whether to extract or abstain. (This contributes up to 0.2 per positive query and up to 1.0 per negative query, i.e., a correctly abstained negative scores the full sample.)
• $R_{\text{company}}$ measures company extraction score, graded credit based on a blend of surface-form F1 and canonical representation Jaccard similarity, plus a small bonus on exact match. (Capped at 0.4 for positive queries and does not contribute for negative queries.)
• $R_{\text{time}}$ measures time-window score, graded credit based on temporal Intersection-over-Union (IoU), plus a small bonus on matching the exact time window. (Capped at 0.4 for positive queries and does not contribute for negative queries.)