diff --git a/BSP.markdown b/BSP.markdown
index 38d1b9e..443b79a 100644
--- a/BSP.markdown
+++ b/BSP.markdown
@@ -14,10 +14,10 @@ Each channel has a unique identifier hash_len bytes long. This identifier is sup
 
 ### Message format
 
-Each message consists of one or more blocks. Each block is block_len bytes long, except the last, which may be shorter. The blocks form the leaves of a binary hash tree, and the root hash of the tree uniquely identifies the message.
+Each message consists of one or more blocks. Each block is block_len bytes long, except the last, which may be shorter. The blocks form the leaves of a binary hash tree. Each internal node of the tree consists of the concatenated hashes of its children, and the hash of the root node uniquely identifies the message.
 
 A message header is prepended to each node of the tree before hashing. The message header consists of the channel identifier, a timestamp indicating when the message was posted, the depth of the hash tree, and the message type.
 
 The timestamp is a 64-bit big-endian integer representing seconds since the Unix epoch. The depth of the hash tree is an 8-bit integer. The message type is a single byte that is not interpreted by BSP; the application may use it to respond appropriately to partially synchronised messages. The total length of the message header is hash_len + 10 bytes.
 
-Prepending the message header to each node before hashing ensures that all blocks with a given message identifer have consistent message headers; this simplifies the validation of blocks received from the remote peer.
\ No newline at end of file
+Prepending the message header to each node before hashing ensures that all blocks with a given message identifer have consistent message headers; this simplifies the validation of blocks received from the remote peer.