[guided]Fix a moment order $m\in\mathbb N$. The object we need to understand is the $m$th free cumulant of $S_n$, where
\begin{align*}
S_n := n^{-1/2}\sum_{i=1}^n a_i.
\end{align*}
Because $\kappa_m:\mathcal A^m\to\mathbb C$ is multilinear in its $m$ arguments, expanding each copy of $S_n$ gives
\begin{align*}
\kappa_m(S_n,\dots,S_n)=n^{-m/2}\sum_{i_1=1}^n\cdots\sum_{i_m=1}^n \kappa_m(a_{i_1},\dots,a_{i_m}).
\end{align*}
The key free-probability input is that mixed cumulants vanish for freely independent families (citing a result not yet in the wiki: vanishing of mixed free cumulants for free subalgebras). Here “mixed” means that the entries of the cumulant come from at least two different free subalgebras. Since the elements $a_i$ are freely independent, the summand
\begin{align*}
\kappa_m(a_{i_1},\dots,a_{i_m})
\end{align*}
is zero unless all indices $i_1,\dots,i_m$ are equal. Thus the full $m$-fold sum collapses to the $n$ diagonal choices
\begin{align*}
(i_1,\dots,i_m)=(j,\dots,j)\quad\text{with }1\leq j\leq n.
\end{align*}
For each such $j$, identical distribution gives
\begin{align*}
\kappa_m(a_j,\dots,a_j)=\kappa_m(a_1,\dots,a_1).
\end{align*}
Consequently,
\begin{align*}
\kappa_m(S_n,\dots,S_n)=n^{-m/2}\sum_{j=1}^n \kappa_m(a_j,\dots,a_j)=n^{1-m/2}\kappa_m(a_1,\dots,a_1).
\end{align*}
This is the scaling mechanism behind the theorem: order $m$ cumulants scale like $n^{1-m/2}$ after normalization by $\sqrt n$.[/guided]