diff --git a/crates/zeta_cli/src/evaluate.rs b/crates/zeta_cli/src/evaluate.rs
index 14dc0f6c0c105919b822b9077211a1e1d9686d04..d808e3d743d7009ca66a75b3a349914b0a4f5447 100644
--- a/crates/zeta_cli/src/evaluate.rs
+++ b/crates/zeta_cli/src/evaluate.rs
@@ -478,7 +478,7 @@ fn evaluate(example: &Example, preds: &PredictionDetails, predict: bool) -> Eval
 
             let scores = Scores::new(&expected, &actual_context_lines);
 
-            false_positive_lines.retain(|line| !actual_context_lines.contains(line));
+            false_positive_lines.retain(|line| !expected.contains(line));
 
             if best_alternative_score
                 .as_ref()