papers/morpheus/adoption/adopt-11-scale.tex

\documentclass[11pt,a4paper]{article}
\usepackage{../shared/luxcover}   % black cover (self-contained TikZ logo)
\usepackage{../shared/proposal}   % \proposalhead helpers and the stake environment
\usepackage[utf8]{inputenc}\usepackage[T1]{fontenc}
\usepackage{amsmath,amssymb}
\usepackage{listings}\usepackage{xcolor}
\input{../shared/lstlang}

\newcommand{\code}[1]{\texttt{\small #1}}
\newcommand{\prov}[1]{{\color{muted}\scriptsize\sffamily[#1]}}
\setlength{\emergencystretch}{3em}

\title{It Does Not Scale Today}
\author{Zach Kelling\\\textit{Lux Industries, Inc.}\\\texttt{[email protected]}}
\date{2026}

\begin{document}
\luxcoverpage
\sloppy

\noindent{\color{muted}\footnotesize\sffamily ADOPT-11 \quad\textbullet\quad Hanzo AI \& Lux Industries \quad\textbullet\quad 2026}\par
\vspace{0.9ex}

\noindent\textbf{The number: a few sessions per second.} \prov{verified}
The hosted gateway at \code{api.mor.org} funds every Web2 caller from one shared consumer
wallet. The happy path runs concurrent. After a nonce conflict on a session open or close,
on-chain ops serialize on the wallet lock for twenty seconds, and each new conflict re-arms it
(\code{config.py:319--323}). One wallet is one nonce sequence, so under contention the gateway
clears a few sessions per second. That is a hard architectural ceiling, not a tuning knob. It is
a load test, not an opinion.

\vspace{0.4ex}
\noindent\textbf{A million users, three walls, in order.}
\begin{itemize}
\item \textbf{Wallet.} One nonce sequence, a few sessions per second \prov{verified}. Users do
      not queue, they collide, and every collision arms the twenty-second lock. There is no
      admission control and no backpressure; the gateway grep is empty \prov{verified}.
\item \textbf{Supply.} Capacity is \code{concurrentSlots} per node, dozens, gated at
      \code{capacity\_manager.go:55}, times 18 providers on \code{active.mor.org}: a few hundred
      concurrent sessions \prov{verified}. Most nodes resell Venice and OpenAI keys under their
      own limits; the node docs show a throttled backend returning \code{429} \prov{verified}.
\item \textbf{Token.} MOR trades at \$1.58, cap near \$14M \prov{verified}. The stipend is 1\% of
      the pool per day, about 29{,}700 MOR \prov{derived}; across a million users that is about
      0.03 MOR each \prov{illustrative}, nothing. So everyone pays direct, and paying direct buys
      MOR through about \$1.3M of DEX depth, mostly protocol-owned, on about \$57k daily volume
      \prov{verified}.
\end{itemize}

\vspace{0.2ex}
\noindent\textbf{The fix, named to a repo.}
Delegated sessions let one funded key authorize many callers, so opens stop serializing on a
single nonce. The Delegation facet is already in \code{Morpheus-Lumerin-Node/smart-contracts}
(\code{Delegation.sol}) \prov{verified}; LP-320 uses it. Enso, Hanzo's router, puts admission
control and backpressure in front of supply \prov{verified}. Settlement moves onto \code{mord},
where a session is per-session state a validator set orders, not a transaction from one key.

\vspace{0.2ex}
\noindent\textbf{The outcome, measured.} \prov{verified: mord/BENCHMARKS.md, M1 Max, single core, in-memory, 2026-08-29}
mord settles the arithmetic in about 2 microseconds per session, roughly 500{,}000 per second on
one core. The un-fakeable cost is two ed25519 verifies per meter, the provider's and the
consumer's, about 296 microseconds, so about 3{,}400 verified sessions per second on one core.
Verification shares no state, so it scales close to linearly: about 25{,}000 per second at eight
cores. The single-wallet wall is gone; the bottleneck becomes how fast consensus orders blocks,
which Lux already answers. Honest boundary: 3{,}400/sec/core is the settlement and verification
rate, not end-to-end network TPS, which consensus and networking set. mord builds, 40 tests pass;
faked work earns zero, a substituted model is slashed \prov{verified}.

\vspace{0.2ex}
\noindent\textbf{Relevance.}
You cannot serve the demand we bring on one wallet and a few hundred slots. The infrastructure has
to go horizontal, and ours is. Keep Morpheus thin: MOR stays a budget and rewards token. Inference
and proof of work done run on A-Chain (\code{lux/ai}, \code{lux/chains/aivm}: bonded operators,
commit-reveal quorum, \code{SettleQuorum}, \code{SlashProvider}, local attestation, no cloud)
\prov{verified}; agentic workloads on the Hanzo agentVM (LP-322 coordination, LP-325 budgets);
post-quantum settlement and consensus on Lux primary; governance on Zoo. Morpheus rebuilds none of
it; it taps in. An agent runs on the agentVM, verifies inference on A-Chain, and settles budget in
MOR.

\vspace{0.6ex}
\noindent{\color{muted}\footnotesize\sffamily
ADOPT-11. Wall at \code{config.py:319--323}; throughput at \code{mord/BENCHMARKS.md}. Fix: LP-320
delegated sessions with Enso admission. Companions ADOPT-01, ADOPT-02, ADOPT-07. Citations checked
against the working tree.\quad\textbullet\quad \texttt{[email protected]}}

\end{document}