Research
Why Ethereum Needs Fairness Mechanisms that Do Not Depend on Participants' Altruism
Ethereum's censorship resistance depends on proposers being willing to include transactions even when excluding them would pay better. Recent fairness mechanisms such as Inclusion Lists (ILs) and…

- arXiv
- 2603.05666
- Published
- 2026-03-05
- Authors
- Patrick Spiesberger, Nils Henrik Beyer, Hannes Hartenstein
AI summary
Why Ethereum Needs Fairness Mechanisms that Do Not Depend on Participants' AltruismOverview
Research area: Ethereum block production, Proposer-Builder Separation (PBS), Maximum Extractable Value (MEV), transaction censorship, and the empirical measurement of validator/proposer behavior.
Technical level: Intermediate. The paper combines protocol-level concepts (MEV-Boost relays, coinbase addresses, priority fees, Inclusion Lists) with a straightforward empirical classification pipeline. Readers familiar with Ethereum staking will follow it easily; beginners will need some background on PBS and MEV.
Scope (one sentence): An empirical study of the full year 2025 that estimates, from observable on-chain and relay data, what share of Ethereum proposers behave altruistically, and what that share implies for committee-based fairness mechanisms that assume at least one honest member per committee.
What This Paper Is About
Ethereum's censorship resistance depends on proposers being willing to include transactions even when excluding them would pay better. Recent fairness mechanisms such as Inclusion Lists (ILs) and Multiple Concurrent Proposer (MCP) designs replace single-proposer control with a committee and assume that at least one member ("1-of-n honest") will refuse to deviate — an assumption usually justified by a guess that a few percent of proposers are altruistic. The paper measures that share empirically instead of assuming it, and asks whether mechanisms built on that assumption remain practical given what proposers actually do.
Key Contributions
- A behavior-based operationalization of altruism for Ethereum proposers. Rather than trying to infer motives, the authors define altruistic behavior as adherence to Ethereum's intents without deviation, and identify deviations along two observable dimensions: MEV extraction or its delegation to external builders, and shared account governance.
- A six-classifier measurement pipeline applied to a full year of data. Starting from 1,108,992 distinct proposers and 2,610,762 blocks (slot range [10,738,799; 13,366,798], skipped slots excluded), the pipeline progressively excludes proposers with unambiguous evidence of non-altruistic behavior, yielding an upper bound ("potentially altruistic") and, with stricter criteria, a lower bound ("strictly altruistic").
- Quantified bounds on altruism prevalence. The share of altruistic proposers is placed in the interval [0.41%; 1.55%] — at most 1.55%, which the authors interpret as an upper bound, with the remaining 98.45% of proposers exhibiting observable non-altruistic behavior.
- An assessment of what these bounds mean for committee-based fairness mechanisms, showing that altruism-dependent designs would need far larger committees than currently proposed, and surveying alternative designs in which fair behavior is the rational choice.
Main Findings
- Relay mediation is near-universal at the block level: 2,383,636 blocks (91.3%) were announced by at least one MEV-Boost relay during the observation period, consistent with prior studies. The relays considered were Aestus, Agnostic, bloXroute (Regulated and Max-Profit), Flashbots, Ultrasound, and Titan.
- Three proposer groups by relay usage: 964,951 proposers (87%) had blocks exclusively announced via relays; 83,150 (7.5%) had a proper subset of blocks announced via relays; 60,891 (5.5%) had blocks (median: two blocks) never announced via relays. Classifier I therefore excludes 94.5% of proposers that used block-building services at least once.
- Governance clustering shrinks the candidate set sharply: applying the shared-governance classifier (Classifier II) reduces the set to 1.59%; adding the builder-payment classifier (Classifier III) reduces the upper bound to 1.55% (17,144 proposers).
- Direct MEV evidence pushes the bound lower: the lenient exclusive-order-flow classifier (Classifier IV) removes no further proposers, leaving 1.55%; the strict version (Classifier V) reduces the set to 1.0%; the strict ordering classifier (Classifier VI) yields 0.41% (4,598 proposers) as the "strictly altruistic" lower bound.
- Excluded proposers are still mostly close to fee-ordered: among proposers excluded by Classifier VI, 99.3% exhibit a Spearman coefficient σ_rs > 0.7 and 35.7% reach σ_rs > 0.99 — suggesting Classifier VI understates genuine altruism.
- The gap versus common assumptions is large: the paper notes the common reading of the 91% figure implies roughly 9% of proposers forgo gains, whereas at most 1.55% can plausibly be regarded as altruistic — almost an order of magnitude below that assumption.
- Committee viability consequences: a 16-member committee (as in the Fork-choice enforced Inclusion Lists design, scheduled for the Hegota Upgrade expected by end of 2026) contains at least one altruist with probability 1 − 0.913^16 ≈ 76.7% under the 8.7%/9% assumption, but only 1 − 0.9845^16 ≈ 22.1% under the measured 1.55% share. Reaching at least 80% probability requires more than 100 members (128 members yield 1 − 0.9845^128 ≈ 86.4%).
- Concentration persists: three clusters, attributable to the two largest builders, account for more than 80% of all blocks; separately, the paper cites that the two largest builders construct over 80% of all blocks and that 88.1%–94.3% of daily blocks were mediated via MEV-Boost relays between Q3 2023 and Q4 2025.
- Quarterly stability of the upper bound: Q1 yields 2.12%, and the share stabilizes within 1.64%–1.68% across Q2, Q3 and Q4, close to the aggregated 1.55%, indicating robustness to observation-window choice. The lower bound varies more across quarters (0.79% in Q1, 0.52% in Q2, 0.63% in Q3, 0.66% in Q4); the provided text is truncated at the discussion of this point.
- Altruistic proposers skew older: 44% of potentially altruistic proposers registered before The Merge (45.9% among strictly altruistic ones); 15% of proposers in validator-index range [0; 100,000) are classified as potentially altruistic versus 0.59% in [1,000,000; 2,000,000).
- Altruistic proposers cluster small: the median cluster among potentially altruistic proposers contains a single proposer, while some reach over 900 members; no cluster of size ≥ 8 was excluded based on only a minority of its proposers.
Methodology in Plain English
The authors start from every proposer that produced at least one block in 2025 and repeatedly remove proposers for whom there is clear evidence that they do not act altruistically. Whatever remains is treated as the candidate altruistic set.
The evidence comes from four sources: a self-operated execution-layer node (transactions, priority fees, coinbase addresses), the Beaconcha.in API for proposer identities, the public endpoints of well-known MEV-Boost relays for which blocks were relayed and who received builder payments, and a network-layer dataset from Yale's Decentralized Systems Lab (transactions observed by four monitoring nodes) to spot transactions that were never publicly propagated, cross-checked against Etherscan.
The six classifiers work like sieves. First, anyone whose block was ever announced by a relay is removed, because those proposers sign blocks they cannot inspect. Second, remaining proposers are grouped by the coinbase address that receives their transaction tips — a proxy for a shared managing party such as a staking provider — and a whole group is removed if any member is already excluded. Third, if a relay reports that a builder's winning bid was paid to an unexcluded group's coinbase address, that group is removed too. Fourth, fourth-and-fifth classifiers look for transactions absent from the public mempool dataset: the lenient version only excludes proposers with more than one such transaction per block (to avoid counting data artifacts), the strict version excludes anyone with at least one. Sixth, the strictest check verifies that transaction ordering is explainable purely by publicly observable priority fees, using a Spearman rank correlation of exactly 1 as the criterion; only individual proposers, not clusters, are excluded here, since failing this test may reflect poor verifiability (for instance user-specified ordering or account-nonce dependencies) rather than misbehavior.
The lenient path yields the upper bound, the strict path the lower bound. Each quarter was also evaluated separately to check temporal stability.
Why This Matters
Impact on research: Prior work on Inclusion Lists and Multiple Concurrent Proposers has largely reasoned about security under an assumed altruism share. This paper supplies an empirical value for that parameter and shows it is roughly an order of magnitude smaller than the commonly cited assumption, which changes the feasibility calculus for a whole family of designs. It also connects to Wadhwa et al.'s result that a 1-of-n honest assumption is fundamentally fragile for order-policy enforcement when all committee members are rational.
Real-world applications:
- Protocol design for FOCIL/Hegota: committee-size parameters must be revisited if inclusion guarantees are to hold at the intended probability.
- Staking provider and operator policy: shared governance structures (self-managed and Staking-as-a-Service, including staking pools) are shown to propagate behavior across many validators, so due diligence must happen at the managing-party level, not the validator level.
- Client configuration choices: settings such as Prysm's
local-block-value-boostcan make a proposer look altruistic on a given block while retaining a profit-maximizing rule. - Designs that align fairness with incentives: reward-based inclusion payments (Fox et al.), AUCIL (Wadhwa et al.), and accountability/disclosure-based builder obligations (Spiesberger et al.) offer alternatives that do not rest on altruism.
Industry relevance: Builder concentration (three clusters accounting for more than 80% of blocks) and near-universal relay usage (91.3% of blocks) mean the practical decision rights over transaction inclusion sit with a very small set of actors. Any mechanism that assumes decentralised, independent proposer judgment needs to account for this.
Future Directions
- Re-evaluate committee sizing under measured altruism: the paper shows 128-member committees reach only about 86.4% inclusion probability at the 1.55% share; open questions include whether such committee sizes are operationally viable and how they interact with block-production latency.
- Measure the cost of bribing committees: the paper states that practically realizing the bribery of committee members into protocol deviations with small incentives "warrants further investigation."
- Account for enshrined PBS (EIP-7732): ePBS was not deployed on Ethereum as of 2025; the authors flag its impact for discussion, but the provided text does not report that analysis.
- Refine the lower bound: Classifier VI excludes proposers whose ordering is not verifiably content-independent even when it may be legitimate, and the authors note this likely understates genuine altruism; better verifiability of ordering (and of user-specified preferences) could tighten the interval. The paper also declines to de-anonymize large clusters for ethical reasons, leaving their identity an open question.
Target Audience
Ethereum protocol researchers and client developers working on inclusion lists, PBS/ePBS, and censorship resistance; staking providers and validator operators interested in how governance structures affect observable behavior; mechanism designers studying incentive-compatible fairness; and data-driven blockchain researchers who want a reusable, reproducible classification pipeline for proposer behavior. Readers seeking a mathematical protocol proof will not find one here — the contribution is empirical measurement plus its design implications.
Authors’ abstract
Ethereum's ideal of censorship resistance, together with related fairness properties, is undermined in practice, motivating fairness mechanisms that aim to restore these properties. Several of these mechanisms hand control over block contents to a committee of proposers under a 1-of-n honest assumption: at least one committee member complies with the mechanism even when deviating would increase personal revenue. We refer to such proposers as altruistic. Yet prior work shows…
The full abstract on arXiv