the total transaction capacity scales with the total number of nodes.

At a certain frequency the software chooses a transaction that references
a high number of parent nodes.

Of course, not every past transaction could be reachable, as users may decide to never reference
certain published transaction. However, the parent selection, with average out-degree 2, and low network
latency, can guarantee that there will be frequent checkpoints referencing almost all previous
transactions.

After a checkpoint is found, the software updates all nodes reachable from the checkpoint with a
forward pointer to this checkpoint. A checkpoint has its own score counter, initially set to zero. When
the update algorithm reaches a checkpoint node, it increments the counter and stops propagating
backward. The score of a transaction is computed as the last stored score in the transaction plus the
score of the pointed checkpoint. Checkpoints are considered as nodes on the DAG, so the same checkpoint
finding algorithm can make checkpoints that refer to other forward checkpoints.