# Multimmit: Extending Blocks for Faster Finality

Crypto.world

> Modern blockchains let many processors disseminate transaction blocks in parallel while a consensus protocol orders small references to that data. The design question is when a block may enter the…

**arXiv**: 2607.21021

**Published**: 2026-07-23

**Authors**: Andrew Lewis-Pye, Patrick O'Grady

## AI summary

## Overview

**Research area:** Distributed computing and blockchain consensus — specifically Byzantine Fault Tolerant State Machine Replication (SMR) in the partially synchronous model, and the design of multi-chain (multi-proposer) consensus protocols.

**Technical level:** Advanced. The paper is a formal protocol design paper with definitions, a complete specification, consistency and liveness proofs, and analysis of extraction thresholds; it assumes familiarity with quorums, view-based BFT, threshold signatures and quality certificates.

**Scope:** The paper introduces Multimmit, a consensus protocol for `n ≥ 5f+1` processors that couples a one-round-of-voting consensus layer with multi-chain data dissemination, and argues it achieves lower latency, better fault isolation and stronger censorship resistance than the certified design Autobahn and the uncertified/prefix-voting design Raptr.

## What This Paper Is About

Modern blockchains let many processors disseminate transaction blocks in parallel while a consensus protocol orders small references to that data. The design question is *when* a block may enter the ordering process: the “certified” approach waits for a quorum attestation of availability (robust but slow), while the “uncertified” approach references data immediately (fast but fragile when data is missing). The paper proposes a third design that keeps data fetch off the critical path, removes the order-sensitivity weakness of the state-of-the-art Raptr protocol (“easy spoiling”), and pushes finality latency down to `t+3δ` in expectation for a transaction block disseminated at time `t`.

## Key Contributions

1. **A new protocol, Multimmit.** A protocol for `n ≥ 5f+1` processors combining multi-chain data dissemination (as in Autobahn) with the consensus skeleton of Minimmit, so that one round of voting per view suffices for finality. Views with correct leaders finalise at `n−f` votes, and every view produces either a V-notarisation or a nullification.
2. **Proposal-relative voting.** The leader issues a *chain proposal* naming tips for every chain (certified or not), and each voter returns a single vote containing one coordinate per chain, reporting how far up each chain it can endorse. A withheld block costs only its own chain’s position; no processor ever blocks or fetches on the critical path.
3. **Extension voting.** Votes may endorse fresh blocks beyond the proposed tips, so a transaction block need only reach the *voters* before they vote rather than the *leader* before it proposes. This removes one message delay from every transaction’s critical path and yields a block-inclusion (censorship-resistance) guarantee that does not depend on the leader.
4. **Full specification and analysis.** A complete implementable specification, proofs of consistency and liveness, an accounting of data availability for every ordered block, exact derivation of the protocol’s extraction thresholds (shown optimal at `n = 5f+1`, with the precise relaxation available for larger `n`), and concrete size calculations for every transmitted object.

## Main Findings

- **Good-case latency:** A transaction block disseminated at time `t` is placed in the total ordering by `t+3δ` in expectation and by `t+2δ` at best, where `δ` is the actual least upper bound on message delay after GST. Crucially, these figures are measured from dissemination of the transaction block itself, not from the leader’s proposal.
- **Comparison with Raptr:** Raptr’s corresponding good-case figure is `t+5δ` in expectation. Of the roughly `2δ` saved, the paper attributes one `δ` to the stronger resilience assumption `n ≥ 5f+1` (one round of voting rather than two) and the remainder to its new mechanisms.
- **Common-case behaviour:** With a correct leader but up to `f` other faulty processors, an honest chain’s block disseminated at `t` is still finalised by `t+3δ` in expectation, and its placement in the total ordering completes by `t+5δ` in expectation in any case. Raptr degrades here: faulty producers can deny it optimistic view exits, stretching views from `2δ` to `3δ`, so the same block is first finalised at `t+5.5δ` in expectation.
- **Easy spoiling has no analogue.** In Raptr, withholding one early batch lets a single faulty producer finalise little or nothing, and `k` faulty producers taking turns deny the optimistic path for `k` consecutive proposals. In Multimmit a faulty producer can significantly delay only its own chain’s blocks; other chains pay at most a one-view wait for placement.
- **Censorship resistance:** If a fresh transaction block from a correct processor reaches all correct processors before they vote in a view, then either the view finalises nothing at all, or the block’s membership in the ledger is settled in that view and it enters the ordering by the next finalised leader block at the latest. A censoring leader must suppress its entire view, forgoing finalisation rewards.
- **Small consensus objects:** Consensus objects total tens of kilobytes per view, independent of transaction volume. Vote vectors condense well — a vote endorsing every proposed tip can be expressed in constant size, and in general need record only deviations from full endorsement.
- **Data-availability accounting:** DA-votes are sent to the block producer alone, who forms a DA-certificate from `n−2f` shares; this costs one extra message delay but avoids cubic communication per block height. A certificate guarantees at least `n−3f ≥ 2f+1` correct holders and uniqueness of certified blocks per chain height.
- **Pipelining:** Producers may run up to `d` blocks ahead of their last certified block (e.g. `d = 3`), with certificates forming in the background — unlike Autobahn, which requires each transaction block to carry its parent’s certificate and thereby caps chain growth.
- **Motivating measurement (from prior work):** A total message loss rate of `0.05%` across the network (one per cent of egress traffic at five per cent of processors) has been observed to increase the latency of Mysticeti, which references uncertified data, by an order of magnitude.
- **Experimental results:** The paper’s Section 7 is described as covering experiments, but the experimental setup, benchmarks and measured numbers are not reported in the excerpt provided.

## Methodology in Plain English

The authors take an existing protocol family (Autobahn-style multi-chain designs, where each processor builds its own chain and a leader-based protocol orders references to those chains) and an existing consensus core (Minimmit, which reaches finality in one round of voting under `n ≥ 5f+1`), and fuse them. They first identify the two failure modes of prior work: Autobahn’s certified approach forces a three-delay round trip (block, DA-votes, certificate) before a block can be referenced, and Raptr’s uncertified prefix voting is fragile to ordering because one withheld batch early in a proposal spoils everything after it. They then design two mechanisms — per-chain proposal-relative votes and extension votes — around a stated design principle: a faulty non-leader should only be able to significantly delay blocks in its own chain. The work proceeds from an informal walkthrough of the design, to a formal specification, to proofs of consistency and liveness, to an exact analysis of extraction thresholds, to size calculations for every object transmitted, with experiments deferred to Section 7 and optimisations (including a backup-proposal extension for leaders that crash before proposing) in Section 6.

## Why This Matters

**Impact on research.** The paper reframes the certified-versus-uncertified debate as a spectrum and shows a point that is simultaneously faster than the certified design and more robust than the uncertified one. It names and formalises “easy spoiling” as a distinct weakness of prefix-voting designs, supplies extraction-threshold optimality results at `n = 5f+1`, and makes the latency accounting explicit about where measurement starts — a methodological point that matters when comparing protocols.

**Real-world applications:**

- High-throughput payment and exchange settlement networks, where user-facing latency is the product.
- Layer-2 rollups and modular blockchain architectures that separate data dissemination from ordering.
- Validator and node infrastructure, where operator API nodes run their own producer chains and place user transactions directly into consensus rather than forwarding them an extra hop.
- Permissioned or consortium chains with a known validator set, where the stronger `n ≥ 5f+1` assumption can be engineered into membership rules.

**Industry relevance.** The paper targets a practical bottleneck: leader bandwidth (hence the erasure-coding work in Kudzu and Alpenglow, which pays an expansion rate typically around three) and the fragility of optimistic designs under realistic loss rates. Its guarantees are framed for *greedy* Byzantine leaders — those that want finalisation but are tempted to censor — which is a realistic adversary model for deployed systems.

## Future Directions

- **Empirical evaluation.** Section 7 describes experiments, but the results are not in the provided excerpt; benchmarking Multimmit against Raptr and Autobahn under loss, asynchrony and Byzantine producers is the natural next step.
- **Deployment and engineering.** Implementing the specification in a production client, including the vector-vote encoding and certificate aggregation across the `n−2f` and `2f+1` threshold schemes.
- **Relaxing the resilience assumption.** The `n ≥ 5f+1` requirement (versus `n ≥ 3f+1` for Autobahn and Raptr) is explicitly flagged as the price of one-round finality; the paper notes the exact relaxation available for larger `n`, leaving open how far the gap can be narrowed in practice.
- **Fault handling and extensions.** The backup-proposal extension (so a leader that crashes before proposing finalises as in a correct-leader view), extension of the analysis to bounded clock-speed differences, and alternatives to Raptr-style reputation mechanisms for admitting previously slow processors.

## Target Audience

Consensus and distributed-systems researchers familiar with BFT SMR and partial synchrony; protocol engineers building or auditing blockchain consensus clients; and technically sophisticated readers who want a rigorous but readable account of how multi-chain dissemination and one-round voting interact. The intuition sections (1–3) are accessible to intermediate readers; the specification and proofs require an advanced background.

## Authors’ abstract

To meet the throughput demands of modern blockchain systems, protocols for State Machine Replication (SMR) increasingly have many processors disseminate blocks of transactions in parallel, with consensus then establishing a total ordering on the blocks of all producers. Such designs face a choice as to when a block may enter the ordering. Certified approaches wait for a quorum to attest a block's availability, which is robust but adds message delays to every transaction.…

[The full abstract on arXiv](https://arxiv.org/abs/2607.21021)

[Read the original paper](https://arxiv.org/abs/2607.21021)

---

Canonical: https://crypto.world/paper/multimmit-extending-blocks-for-faster-finality-2607.21021
