mirror of
https://github.com/zoo-apps/papers.git
synced 2026-11-02 02:29:15 +00:00
zeekay
95d7b38d07
The execution-proof slice the paper marked "designed, not built" is now built and tested; update the paper to reflect the verified reality and add the Zoo L3 settlement / DeSci framing the paper was missing. Grounded in code read + tests run for this paper: - Rust engine (hanzo-engine poi_*): 27 unit + 11 adversarial = 38 passing (poi.rs 7, poi_transcript.rs 8, poi_forward.rs 3 ProvableLinear, poi_graph.rs 9 GraphTrace incl. a training step). Cross-language golden vector pins Go parity. - Go verifier (luxfi/crypto/poi): 23 passing (TestDeriveChallengesGolden parity). - Solidity (luxfi/standard contracts/ai/thinking): 87 passing across the 8 PoAI suites (ComputeWitnessLib 5, OptimisticEvidence 18, AICoinMiner 12, ComputeProofEnforcement 19, AICoin 11, AIComputeRegistry 6, MinerStakeRegistry 8, L3Registry 8). AICoinMiner mint gate fails closed (ProofGateUnavailable). - Confidentiality primitives: crypto/promptseal (X-Wing) 6, crypto/attestation 9. Changes: - New section 18 (24-zoo-l3-settlement): the Zoo L3 consumes the one canonical PoAI contract surface; L3Registry directory; trustless ML-DSA validator-quorum root relay (AChainRootOracle); fail-closed mint + global no-double-mint; and proven inference/embedding/training as the DeSci reproducibility substrate. - Rewrote status-summary (00) and implementation-status (23) as a verified ledger with real passing counts; flipped the stale "designed, not built" framing. - Abstract/intro/conclusion/future-work updated to the built reality; residual honestly scoped to the RepOps fp tier + chaining/long-tail emission. - Reconciled stale "0% / no caller / once emission exists / roadmap" passages in sections 11, 16, 17, 18, 19, 20, 21, 22; experimental numbers marked as illustrative projections (not measured), with the measured asymmetry cited. - Cite the zoo/proofs soundness + emission companions. LaTeX compiles clean: pdflatex x3, 52 pages, 0 undefined refs/citations.
110 lines
3.5 KiB
TeX
110 lines
3.5 KiB
TeX
\section{Experimental Evaluation}
|
|
|
|
\label{sec:evaluation}
|
|
|
|
\begin{quote}\small\itshape
|
|
Status note. The throughput, overhead, and detection figures in this section are
|
|
\emph{illustrative projections} of the consensus design under the stated configuration,
|
|
not measurements from a running testnet; we mark them as such so no reader mistakes a
|
|
target for an observed result, in keeping with the up-front-honesty discipline of
|
|
Section~\ref{sec:status-summary}. The figures that \emph{are} measured live elsewhere in
|
|
this paper and its companions: the execution-proof test suites
|
|
(Section~\ref{sec:impl-status}: $38$ Rust, $23$ Go, $87$ Solidity tests passing, run for
|
|
this paper) and the single-node verification latency of a challenged matmul slice reported
|
|
in the emission companion~\cite{poiemission}. The asymmetry those establish---verifying a
|
|
challenged matmul is an order cheaper than recomputing it (Section~\ref{sec:freivalds})---is
|
|
the load-bearing performance claim; the consensus-level numbers below are the design's
|
|
intended operating point.
|
|
\end{quote}
|
|
|
|
\subsection{Testnet Configuration}
|
|
|
|
\begin{itemize}
|
|
\item \textbf{Nodes}: 100 validators, 500 workers
|
|
\item \textbf{Hardware}: Mix of Intel SGX and AMD SEV-SNP
|
|
\item \textbf{Network}: Geographically distributed (5 continents)
|
|
\item \textbf{Workloads}: Inference (70\%), training (20\%), extraction (10\%)
|
|
\end{itemize}
|
|
|
|
\subsection{Throughput and Latency}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lccc}
|
|
\toprule
|
|
\textbf{Metric} & \textbf{PoAI} & \textbf{Tendermint} & \textbf{PoW (Ethereum)} \\
|
|
\midrule
|
|
Throughput (TPS) & 847 & 1,000 & 15 \\
|
|
Finality (seconds) & 2.3 & 1.0 & 360 \\
|
|
Energy (kWh/tx) & 0.0002 & 0.0001 & 50 \\
|
|
Useful work (\%) & 98\% & 0\% & 0\% \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Consensus comparison. \PoAI{} achieves competitive throughput while performing useful AI work.}
|
|
\label{tab:consensus}
|
|
\end{table}
|
|
|
|
\subsection{Verification Overhead}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lcccc}
|
|
\toprule
|
|
\textbf{Task Type} & \textbf{Compute (ms)} & \textbf{Verify (ms)} & \textbf{Overhead} \\
|
|
\midrule
|
|
Inference (7B) & 150 & 12 & 8.0\% \\
|
|
Inference (70B) & 2,400 & 45 & 1.9\% \\
|
|
Training step & 850 & 23 & 2.7\% \\
|
|
Extraction & 1,200 & 34 & 2.8\% \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Verification overhead. \TEE{} attestation adds minimal latency.}
|
|
\label{tab:overhead}
|
|
\end{table}
|
|
|
|
\subsection{Quality Scoring Accuracy}
|
|
|
|
\begin{table}[t]
|
|
\centering
|
|
\begin{tabular}{lcc}
|
|
\toprule
|
|
\textbf{Attack Type} & \textbf{Detection Rate} & \textbf{False Positive} \\
|
|
\midrule
|
|
Random outputs & 100\% & 0\% \\
|
|
Stale outputs (replay) & 98.7\% & 0.3\% \\
|
|
Sybil quality voting & 99.2\% & 0.5\% \\
|
|
\bottomrule
|
|
\end{tabular}
|
|
\caption{Attack detection. Multi-party scoring resists manipulation.}
|
|
\label{tab:attacks}
|
|
\end{table}
|
|
|
|
\subsection{Reward Distribution}
|
|
|
|
\begin{figure}[t]
|
|
\centering
|
|
\begin{tikzpicture}[scale=0.8]
|
|
\begin{axis}[
|
|
xlabel={Quality Score},
|
|
ylabel={Reward Share (\%)},
|
|
xmin=0, xmax=1,
|
|
ymin=0, ymax=5,
|
|
legend pos=north west,
|
|
grid=major
|
|
]
|
|
\addplot[blue, thick, mark=*] coordinates {
|
|
(0.1, 0.2) (0.3, 0.8) (0.5, 1.5) (0.7, 2.5) (0.9, 4.2)
|
|
};
|
|
\addlegendentry{PoAI}
|
|
|
|
\addplot[red, thick, dashed] coordinates {
|
|
(0.1, 2.0) (0.3, 2.0) (0.5, 2.0) (0.7, 2.0) (0.9, 2.0)
|
|
};
|
|
\addlegendentry{Equal distribution}
|
|
\end{axis}
|
|
\end{tikzpicture}
|
|
\caption{Reward distribution by quality. \PoAI{} incentivizes high-quality work.}
|
|
\label{fig:rewards}
|
|
\end{figure}
|
|
|