diff --git a/src/structs/branch.rs b/src/structs/branch.rs
index ebae81bb..070e5832 100644
--- a/src/structs/branch.rs
+++ b/src/structs/branch.rs
@@ -79,14 +79,8 @@ impl Branch {
             // non-overflow case is easy
             self.counts += 1;
         } else {
-            // special case where it is all trues
-            if self.counts <= 0x01ff {
-                // corner case since the original implementation
-                // insists on setting the probabily to zero,
-                // although the probability calculation would
-                // return 1.
-                self.counts = 0x00ff;
-            } else {
+            // normalize, except special case where it is all trues
+            if self.counts != 0x01ff {
                 self.counts = (((self.counts as u32 + 0x100) >> 1) & 0xff00) as u16 | 129;
             }
         }
@@ -94,17 +88,11 @@ impl Branch {
 
     #[inline(always)]
     pub fn record_and_update_false_obs(&mut self) {
-        if self.counts == 0x00ff {
-            // handle corner case where prob was set to zero (purely for compatibility, remove this if there is a breaking change in the format)
-            self.counts = 0x02ff;
-            return;
-        }
-
         let (result, overflow) = self.counts.overflowing_add(0x100);
         if !overflow {
             self.counts = result;
         } else {
-            // special case where it is all falses
+            // normalize, except in special case where it is all falses
             if self.counts != 0xff01 {
                 self.counts = ((1 + (self.counts & 0xff) as u32) >> 1) as u16 | 0x8100;
             }
@@ -189,7 +177,15 @@ fn test_all_probabilities() {
             old_t.record_obs_and_update(true);
             new_t.record_and_update_true_obs();
 
-            assert_eq!(old_t.probability, new_t.get_probability());
+            if old_t.probability == 0 {
+                // there is a change of behavior here compared to the C++ version,
+                // but because of the way split is calculated it doesn't result in an
+                // overall change in the way that encoding is done, but it does simplify
+                // one of the corner cases.
+                assert_eq!(new_t.get_probability(), 1);
+            } else {
+                assert_eq!(old_t.probability, new_t.get_probability());
+            }
         }
     }
 }