mirror of
https://github.com/zoo-apps/papers.git
synced 2026-10-30 14:56:54 +00:00
zeekay
45cec7a0c7
Map the Thinking*/ProofOfThought*/AICoinMiner contract, interface, and type
names to the AI* public-protocol namespace. The phrase "Thinking Chains"
survives ONLY as the loose name for the category of blockchains/networks that
run this stack, never as a contract/protocol/type name.
- ThinkingGovernor -> AIGovernor
- ThinkingParameters -> AIParams
- ThinkingReputation -> AIReputation
- ThinkingChainObservatory -> AIObservatory
- ProofOfThoughtRegistry -> AIThoughtRegistry
- AICoinMiner -> AIMiner (.sol / .mine / .t.sol / _verifyMerkle)
- cited path contracts/ai/thinking/ ->
contracts/ai/{governance,mining,token,compute,relay,registry}/
Unchanged: AICoin, AIExecute, AIApproval, AIComputeRegistry, AChainRootOracle,
MinerStakeRegistry, ComputeWitnessLib, OptimisticEvidence, L3Registry.
No soundness math, theorems, test counts, or technical content changed.
Recompiles clean (52 pages).
58 lines
2.4 KiB
TeX
58 lines
2.4 KiB
TeX
\section{Security Analysis}
|
|
|
|
\label{sec:security}
|
|
|
|
\subsection{Byzantine Fault Tolerance}
|
|
|
|
\begin{theorem}[Safety]
|
|
If at most $f < n/3$ validators are Byzantine, no two honest validators finalize conflicting blocks.
|
|
\end{theorem}
|
|
|
|
\begin{proof}
|
|
Block finalization requires $2f+1$ valid signatures. With $n = 3f+1$ total validators and at most $f$ Byzantine, any two sets of $2f+1$ signers overlap in at least one honest validator. An honest validator signs at most one block per height.
|
|
\end{proof}
|
|
|
|
\begin{theorem}[Liveness]
|
|
In partial synchrony, if at most $f < n/3$ validators are Byzantine, honest transactions are eventually included.
|
|
\end{theorem}
|
|
|
|
\begin{proof}
|
|
After GST, message delays are bounded. With $2f+1$ honest validators, quorum can always be formed. Leader rotation ensures eventual honest leader.
|
|
\end{proof}
|
|
|
|
\subsection{Incentive Compatibility}
|
|
|
|
\begin{theorem}[Truthful Quality Reporting]
|
|
Under the multi-party evaluation protocol, reporting true quality is a dominant strategy.
|
|
\end{theorem}
|
|
|
|
\begin{proof}
|
|
The trimmed mean ignores extreme values. Lying increases variance but cannot shift the result beyond honest range. Deviating validators are identified and slashed.
|
|
\end{proof}
|
|
|
|
\begin{theorem}[No Free Riding]
|
|
Workers cannot claim rewards without performing computation.
|
|
\end{theorem}
|
|
|
|
\begin{proof}
|
|
\TEE{} attestation binds output to enclave execution, and the software-only execution proof
|
|
binds reward to genuine work with no hardware-trust assumption: a worker that claims an
|
|
output it did not compute commits an activation-trace whose matmul is caught by the
|
|
Freivalds challenge except with probability $\le 2^{-122}$ (Section~\ref{sec:freivalds};
|
|
built and tested, Section~\ref{sec:impl-status}), and the on-chain mint gate fails closed
|
|
without a valid proof (\texttt{AIMiner}, Section~\ref{sec:l3-settlement}). Probabilistic
|
|
challenge catches cheaters with high probability and slashing the bond exceeds the expected
|
|
gain. The execution proof is what makes this hold for an untrusted, hardware-free prover,
|
|
not only inside a trusted enclave.
|
|
\end{proof}
|
|
|
|
\subsection{Sybil Resistance}
|
|
|
|
Sybil attacks are mitigated through:
|
|
\begin{enumerate}
|
|
\item \textbf{Stake Requirement}: Validators must stake tokens
|
|
\item \textbf{Work Requirement}: Workers must perform verified computation
|
|
\item \textbf{Hardware Binding}: \TEE{} attestations bound to physical devices
|
|
\end{enumerate}
|
|
|