CTMCs are normally defined with no self loops. That is, when in state $i$, you transition to state $j$ at a rate $Q_{ij}$. You never transition from $i$ to $i$, so we're free to choose $Q_{ii}$ such that the rows of $Q$ sum to zero, so that probability is conserved (as answered here).
But what if we did allow self loops? I'm designing a simulation where each state $j$ provides a score $S_j \geq 0$, where higher scores make it more likely to transition to that state. But I'd also like high scores to make you more likely to stay in that state. So generally I want to start with $Q'_{ij}=S_j$, but then normalize $Q'_{ij}$ to $Q_{ij}$, with rows summing to zero, without discarding $S_i=Q'_{ii}$.
Intuitively, given a CTMC with self loops on state $i$ (taken at rate $S_i$), I want to find an equivalent one without self loops. It seems that a node with a self loop could be replaced by one without, just with lower outgoing rates. But how exactly to do this?
Since I'm simulating this in discrete time steps anyway, I'm able to simply renormalize the probability distribution before I sample it:
$$P_i(t+{\Delta}t) = \frac{P_i(t)+S_i{\Delta}t}{1+{\Delta}t\sum_{j}S_j}$$
But I'm not having any luck writing that same distribution in terms a $Q$ matrix based on $S$ such that
$$ P_j(t+{\Delta}t) = P_j(t)+Q_{ij}{\Delta}t $$ $$ Q_{ij} \propto S_j, \text{ for all } j \neq i $$ $$ \sum_{j}Q_{ij} = 0 $$