diff --git a/netam/dcsm.py b/netam/dcsm.py
index 5d61bba4..7f4676b2 100644
--- a/netam/dcsm.py
+++ b/netam/dcsm.py
@@ -298,6 +298,7 @@ def predictions_of_batch(self, batch):
         So, in cases when the sum of the children is > 1, we don't give a
         normalized probability distribution, but that's OK for loss calculation
         because we are doing softmax.
+        TODO think about this more.
 
         Note that make all ambiguous codons nan in the output, ensuring that
         they must get properly masked downstream.