Benjamin Marsh

Alejandro Ranchal-Pedrosa, Benjamin Marsh

The TRAP protocol solves rational agreement by combining accountable consensus with a one-shot BFTCR finalization phase. We present SNARE (Scalable Nash Agreement via Reward and Exclusion), the adaptation of TRAP to $n=5f{+}1$, and prove $ε$-$(k,t)$-robustness for rational agreement tolerating coalitions up to ${\approx}73\%$ with deposits under $0.5\%$ of the gain. A central finding is that appending a single all-to-all broadcast round with the $4f{+}1$ threshold after predecisions yields $ε$-$(k,t)$-robustness for coalitions up to $3f$ (${\approx}60\%$) without any deposit: we need not model or know the utility function of deviating players, only that they participate in the protocol. These players can be \emph{deceitful} (arbitrary unknown utility), not just rational, and the finalization structure prevents disagreement regardless of their motivation. This observation is protocol-agnostic, applies to any $5f{+}1$ protocol at the cost of one message delay that runs concurrently with the next view, and does not require commit-reveal mechanisms. Above $60\%$, the full baiting mechanism with deposits under $0.5\%$ extends tolerance to ${\approx}73\%$. A second finding is that valid-candidacy, the property preventing reward front-running, holds unconditionally regardless of the quorum threshold, removing both the $n>2(k{+}t)$ and $n>\frac{3}{2}k{+}3t$ constraints from the original TRAP. This retroactively extends the $3f{+}1$ bound from $C<n/2$ to $C<5n/9$. The binding constraint in both models is the winner consensus operating on $2f$ residual players after excluding $3f{+}1$ detected equivocators. We explore avenues for relaxing this limit.

Benjamin Marsh, Alejandro Ranchal-Pedrosa

Sedna is a coded multi-proposer consensus protocol in which a sender shards a transaction payload into rateless symbols and disseminates them across parallel proposer lanes, providing high throughput and ``until decode'' privacy. This paper studies a sharp incentive failure in such systems. A cartel of lane proposers can withhold the bundles addressed to its lanes, slowing the chain's symbol accumulation while privately pooling the missing symbols. Because finalized symbols become public, the cartel's multi-slot information lead is governed by a chain level delay event where the chain fails to accumulate the $κ$ bundles needed for decoding by the honest horizon $t^\star=\lceil κ/m\rceil$. We characterize the resulting delay probability with KL-type large deviation bounds and show a knife edge pathology when the slack $Δ=t^\star m-κ$ is zero such that withholding a single bundle suffices to push inclusion into the next slot with high probability. We propose \textsf{PIVOT-$K$}, a Sedna native pivotal bundle bounty that concentrates rewards on the $κ$ bundles that actually trigger decoding, and we derive explicit incentive compatibility conditions against partial and coalition deviations. We further show that an adaptive sender ``ratchet'' that excludes lanes whose tickets were not redeemed collapses multi-slot withholding into a first slot deficit when $t^\star\ge 2$, reducing the required bounty by orders of magnitude. We close by bounding irreducible within slot decode races and providing parameter guidance and numerical illustrations. Our results show that for realistic parameters Sedna can reduce MEV costs to 0.04\% of the transaction value.

Bernardo Magri, Benjamin Marsh, Paul Gebheim

Modern cloud inference creates a two sided privacy problem where users reveal sensitive inputs to providers, while providers must execute proprietary model weights inside potentially leaky execution environments. Fully homomorphic encryption (FHE) offers cryptographic guarantees but remains prohibitively expensive for modern architectures. We argue that progress requires co-design where specializing FHE schemes/compilers for the static structure of inference circuits, while simultaneously constraining inference architectures to reduce dominant homomorphic cost drivers. We outline a meet in the middle agenda and concrete optimization targets on both axes.