papers/lux-consensus-engines/sections/08-composition.tex

\chapter{Composition in the Quasar Stack}
\label{ch:composition}

The seven engines are not alternatives---they are layers. The Lux primary
network composes them into a single finality pipeline that produces
\emph{QuasarCerts} (LP-020). This chapter records the composition,
not as new mechanism, but as the contract each engine fulfils for the
others.

\section{Layered Picture}

\begin{center}
\begin{tabular}{p{2.6cm}p{11cm}}
\toprule
\textbf{Layer} & \textbf{Role}\\
\midrule
Wave   & primitive: per-vertex confidence counters, monotone in expectation;
         supplies the safety theorem that every higher layer reuses\\
Photon & pipelining of Wave queries; produces sustained per-conflict
         throughput by overlapping $\beta$-rounds across many in-flight
         transactions\\
Nova   & lifts Photon from individual vertices to DAG cones via the
         supernova burst: a single frontier finalization closes its
         entire ancestor cone\\
Nebula & swaps Wave's i.i.d.\ sampling for stratified sampling with
         bridge validators when the network graph is sparse, lossy, or
         partitioned\\
Prism  & parallelises Photon across $B$ spectrum bands, hashing each
         conflict set into one band; aggregate throughput scales linearly
         in $B$\\
Ray    & feedback layer above Wave/Photon: monitors confidence velocity,
         estimates adversary fraction, and adapts $\beta$ to minimise
         latency at fixed safety\\
Field  & generalisation to $d > 2$ colours through an electromagnetic
         potential; used for validator-set elections, governance and
         multi-party auctions where the decision alphabet is not binary\\
\bottomrule
\end{tabular}
\end{center}

\section{Quasar Pipeline}

A QuasarCert (LP-020) is the artefact of running these engines back to
back on the primary network:

\begin{enumerate}[nosep]
\item \textbf{Photon} pipelines candidate vertices into the per-conflict
sampler.
\item \textbf{Wave} drives those samples until each candidate becomes
strongly preferred (its DAG cone is internally consistent).
\item \textbf{Nova} closes the cone in a supernova burst, finalising
every ancestor of the frontier vertex simultaneously.
\item \textbf{Ray} watches the empirical confidence velocity and
adjusts the global $\beta$ down toward 3 rounds when the adversary
appears benign.
\item \textbf{Nebula} replaces the i.i.d.\ sampler with stratified
sampling when the live validator set is below the expansion threshold
(rare on mainnet, common on appchains).
\item \textbf{Prism} (off by default on mainnet, default on Z-Chain and
B-Chain) partitions the mempool into $B$ bands so steps 1--5 run $B$
times in parallel.
\item \textbf{Field} runs orthogonally for non-binary decisions such as
validator-set epoch elections and DAO votes.
\end{enumerate}

\section{Interfaces and Contracts}

Each engine exposes the same minimal interface:

\begin{itemize}[nosep]
\item \textbf{Submit}$(t)$ -- accept a candidate (vertex, band, or colour).
\item \textbf{Vote}$(t, c)$ -- broadcast the local preference for $t$.
\item \textbf{Sample}$(k)$ -- sample $k$ peers and return their votes.
\item \textbf{Finalize}$(t, c)$ -- declare $t$ irreversibly committed to
colour $c$.
\end{itemize}

This interface is the entire reason composition is possible: the higher
layer treats the lower layer as a black box with these four methods,
and the lower layer treats its own samples as a black-box oracle.
Confidence monotonicity (Wave) is preserved by every wrapper because
each wrapper only adds positive evidence; it never subtracts.

\section{Mainnet Activation}

The composition described here is what activated on the Lux primary
network on 25 December 2025 as Quasar 3.0 (LP-020). Photon, Wave, Nova,
Ray and Field run on every validator. Nebula activates per-chain when
the in-chain validator count drops below the expansion threshold.
Prism is enabled by default on Z-Chain (proof-rollup) and B-Chain
(bridge), where independent conflict sets dominate.