diff --git a/crates/edit_prediction_cli/src/reversal_tracking.rs b/crates/edit_prediction_cli/src/reversal_tracking.rs
index a014539650748188e19e2c3c8a687ed3a4b0ac6f..6f1be51c79df3d002004a09c5d0977d5f2a215eb 100644
--- a/crates/edit_prediction_cli/src/reversal_tracking.rs
+++ b/crates/edit_prediction_cli/src/reversal_tracking.rs
@@ -115,16 +115,22 @@ fn is_subsequence(needle: &str, haystack: &str) -> bool {
     needle_chars.peek().is_none()
 }
 
-/// Normalize edits where `old_text` appears as a subsequence within `new_text`.
-/// When the user's text is preserved (in order) within the prediction, we only
-/// count the newly inserted characters, not the preserved ones.
+/// Normalize edits where `old_text` appears as a subsequence within `new_text` (extension),
+/// or where `new_text` appears as a subsequence within `old_text` (reduction).
+///
+/// For extensions: when the user's text is preserved (in order) within the prediction,
+/// we only count the newly inserted characters, not the preserved ones.
 /// E.g., "epr" → "eprintln!()" becomes 8 inserted chars ("intln!()")
 /// E.g., "test_my_function" → "a_test_for_my_special_function_plz" becomes 18 inserted chars
+///
+/// For reductions: when the prediction's text is preserved (in order) within the original,
+/// we only count the deleted characters, not the preserved ones.
+/// E.g., "ifrom" → "from" becomes 1 deleted char ("i")
 fn normalize_extension_edits(edits: Vec<GranularEdit>) -> Vec<GranularEdit> {
     edits
         .into_iter()
         .map(|edit| {
-            if edit.old_text.is_empty() {
+            if edit.old_text.is_empty() || edit.new_text.is_empty() {
                 return edit;
             }
 
@@ -135,6 +141,13 @@ fn normalize_extension_edits(edits: Vec<GranularEdit>) -> Vec<GranularEdit> {
                     old_text: String::new(),
                     new_text: edit.new_text.chars().take(inserted_len).collect(),
                 }
+            } else if is_subsequence(&edit.new_text, &edit.old_text) {
+                let deleted_len = edit.old_text.len() - edit.new_text.len();
+                GranularEdit {
+                    range: edit.range.start..edit.range.start + deleted_len,
+                    old_text: edit.old_text.chars().take(deleted_len).collect(),
+                    new_text: String::new(),
+                }
             } else {
                 edit
             }
@@ -147,7 +160,8 @@ fn compute_reversal_overlap(
     current_content: &str,
     predicted_content: &str,
 ) -> ReversalOverlap {
-    let history_edits = compute_granular_edits(original_content, current_content);
+    let history_edits =
+        normalize_extension_edits(compute_granular_edits(original_content, current_content));
     let prediction_edits =
         normalize_extension_edits(compute_granular_edits(current_content, predicted_content));
 
@@ -483,6 +497,33 @@ mod tests {
                 expected_reversal_chars: 7,
                 expected_total_chars: 7,
             },
+            Case {
+                name: "fixing typo is not reversal",
+                original: "",
+                current: "<dv",
+                predicted: "<div>",
+                expected_reversal_chars: 0,
+                expected_total_chars: 2,
+            },
+            Case {
+                name: "infix insertion not reversal",
+                original: "from my_project import Foo\n",
+                current: "ifrom my_project import Foo\n",
+                predicted: indoc::indoc! {"
+                    import
+                    from my_project import Foo
+                "},
+                expected_reversal_chars: 0,
+                expected_total_chars: 6,
+            },
+            Case {
+                name: "non-word based reversal",
+                original: "from",
+                current: "ifrom",
+                predicted: "from",
+                expected_reversal_chars: 1,
+                expected_total_chars: 1,
+            },
         ];
 
         for case in &cases {