papers/lux-ray-protocol/sections/06-empirical-evaluation.tex

\section{Empirical Evaluation}

\label{sec:evaluation}

\subsection{Setup}

We deployed Ray on a 1,000-node testnet across 8 geographic regions with parameters
$k = 20$, $\alpha = 0.8$, $\beta_{\min} = 3$, and $\lambda$ calibrated for $2^{-128}$ safety.

\subsection{Latency Results}

\begin{table}[h]
\centering
\begin{tabular}{@{}lrrr@{}}
\toprule
\textbf{Byzantine \%} & \textbf{Median (ms)} & \textbf{P99 (ms)} & \textbf{Improvement vs. Static} \\
\midrule
0\% & 200 & 300 & 64\% \\
10\% & 350 & 600 & 38\% \\
20\% & 500 & 900 & 23\% \\
30\% & 650 & 1200 & 7\% \\
\bottomrule
\end{tabular}
\caption{Ray latency vs. static threshold ($\beta = 14$)}
\label{tab:latency}
\end{table}

Under zero Byzantine faults, Ray achieved 200ms median latency (4 rounds at 50ms),
a 64\% improvement over the static threshold of $\beta = 14$ (700ms). Under 20\%
Byzantine, the improvement was 23\% (500ms vs. 650ms), as the adaptive threshold
correctly increased to account for adversarial interference.

\subsection{Estimator Accuracy}

The adversary estimator $\hat{\gamma}$ converged to within 2\% of the true $\gamma$
after 15 rounds (750ms), with a 95\% confidence interval width of $\pm 3\%$.

\subsection{Throughput}

Ray achieved 4,200 TPS on the 1,000-node testnet, comparable to the static-threshold
protocol (4,100 TPS). The adaptive computation adds negligible overhead (one division
and comparison per round per transaction).

%% -----------------------------------------------------------------------
%%  7. RELATED WORK
%% -----------------------------------------------------------------------