Fault Tolerance and Information Aggregation in Blockchain Consensus

How to cite

Takayama, Shino; Mezzetti, Claudio; Liu, Yuqing; Liu, Yunlong. 2026. "Fault Tolerance and Information Aggregation in Blockchain Consensus." AEA RCT Registry. AEARCTR-0019221. https://doi.org/10.1257/rct.19221-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Shino Takayama* , Claudio Mezzetti , Yuqing Liu , Yunlong Liu

* Primary investigator

Registry ID
AEARCTR-0019221
DOI
https://doi.org/10.1257/rct.19221-1.0
First registered
Last updated
Status
In development
Trial dates
2026-08-03 – 2027-09-30
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/19221

Abstract

This study uses a laboratory experiment to test equilibrium predictions of a theoretical model of blockchain consensus committee voting (Liu et al., 2026). Participants act as members of a three-person committee voting by majority rule to validate or reject blocks under two protocols (sequential vs. simultaneous) and two adversary conditions (q=0 vs. q=0.1). Under the sequential protocol, participants observe prior votes before casting their own; under the simultaneous protocol, they vote without observing others. An adversary replaces committee members' votes with probability q=0.1 in the adversary treatment. The experiment uses a between-subjects design across protocols and a within-subjects design across adversary conditions. Five primary hypotheses test directional predictions about validation error rates derived from the theoretical model. A total of 222 participants (8 sessions) will be recruited from the University of Queensland undergraduate population. The sequential arm will comprise six sessions of 30 participants (180 total), and the simultaneous arm will comprise two sessions of 21 participants (42 total).

Experimental design

Design overview
The experiment uses a 2x2 design crossing protocol type (sequential vs. simultaneous) with adversary probability (q=0 vs. q=0.1). Protocol assignment is between-subjects: participants are assigned to either the sequential or simultaneous protocol for their entire session. Adversary condition is within-subjects: in half of the sequential sessions, rounds 1-20 use q=0 (no adversary) and rounds 21-40 use q=0.1 (possible adversary); in the other half of the sequential sessions, the order is reversed (rounds 1-20 use q=0.1 and rounds 21-40 use q=0). This counterbalancing controls for order effects in the sequential arm. Within each protocol arm, the adversary order is counterbalanced across sessions. In the simultaneous arm, one session uses q=0 followed by q=0.1, and the other uses q=0.1 followed by q=0.

Task and signal structure
Participants act as members of a three-person blockchain validation committee. In each round, the committee votes by majority rule to validate (Yes) or reject (No) a block. The block is legitimate (good state, theta=G) or illegitimate (bad state, theta=B) with equal prior probability 1/2. Each participant receives a private binary signal: white (good signal) or black (bad signal), where Pr(white|G)=1 and Pr(white|B)=1/2. In the sequential protocol, participants vote one at a time and observe prior votes. In the simultaneous protocol, all participants vote without observing others.

Adversary treatment
In q=0.1 rounds, each participant's vote is independently replaced by an adversary's vote with probability 10%. The adversary votes against the socially correct outcome. Participants are informed of this probability but do not observe whether replacement occurred until the end-of-round results screen.

Payoffs
+1 point if the participant voted Yes, was not replaced, and the group correctly validated a legitimate block; -1 point if the participant voted Yes, was not replaced, and the group incorrectly validated an illegitimate block; 0 otherwise. Points are converted to cash at $0.50 per point.

Sessions and randomisation
Sessions are conducted in the UQ experimental economics laboratory using oTree software. Within each session, participants are randomly re-matched into groups of three each round. In the sequential protocol, voting order is randomly assigned each round. Participants are assigned to one of eight sessions. The six sequential sessions each enrol 30 participants, while the two simultaneous sessions each enrol 21 participants. Within each protocol arm, the adversary order is counterbalanced across sessions.

Primary hypotheses (all two-sided at alpha=0.05, Bonferroni-corrected threshold alpha*=0.010)
• H1a: Pr(approved|B, Simultaneous, q=0) > Pr(approved|B, Sequential, q=0)
• H1b: Pr(approved|B, Simultaneous, q=0.1) > Pr(approved|B, Sequential, q=0.1)
• H2: Pr(rejected|G, Sequential, q=0.1) > Pr(rejected|G, Simultaneous, q=0.1)
• H3a: Pr(approved|B, Sequential, q=0.1) > Pr(approved|B, Sequential, q=0)
• H3b: Pr(rejected|G, Sequential, q=0.1) > Pr(rejected|G, Sequential, q=0)
Statistical analysis
Primary statistical tests are two-sample z-tests for proportions at the group-round level. A linear probability model (LPM) is estimated as the primary regression, with probit as a robustness check. A Fisher combined probability test across all five hypotheses (k=5) serves as a global test of the theoretical model, achieving greater than 99% power at the planned sample size.

Design details

Randomization method
Participants are randomly assigned by computer to either the sequential or simultaneous protocol at the session level. Within each session, participants are randomly re-matched into groups of three at the beginning of every round. In the sequential protocol, the voting order within each group is randomly determined by the oTree software in every round. Within each protocol arm, sessions are allocated so that half follow the q = 0 → q = 0.1 order and half follow the q = 0.1 → q = 0 order, thereby counterbalancing order effects. In q = 0.1 rounds, each participant's vote is independently replaced by the experimental software with probability 0.1.
Randomization unit
Randomization occurs at multiple levels. Participants are randomized to experimental sessions, and sessions are assigned to one of the two protocol treatments (sequential or simultaneous). Within each session, participants are randomly re-matched into three-person groups in every round. In the sequential protocol, voting order is randomized within each group in every round. In rounds with adversary probability q = 0.1, adversary replacement is independently randomized at the individual participant level.
Clusters
8 sessions.
Observations
222 Participants in total
Arms
• Sequential arm: 6 sessions x 30 participants = 180 participants • Simultaneous arm: 2 sessions x 21 participants = 42 participants Approximately 8,880 individual decisions and 2,960 group-rounds in total, comprising 2,400 group-rounds in the sequential arm (6 sessions × 40 rounds × 10 groups) and 560 group-rounds in the simultaneous arm (2 sessions × 40 rounds × 7 groups).
Primary outcomes
1. Illegitimate block approval rate: the proportion of group-rounds in the bad state (theta=B) in which the group decision is approved (majority votes Yes), separately for q=0 and q=0.1 and each protocol. 2. Legitimate block rejection rate: the proportion of group-rounds in the good state (theta=G) in which the group decision is rejected (majority votes No), separately for q=0 and q=0.1 and each protocol.
Secondary outcomes
3. Individual vote compliance: proportion of benevolent participants voting correctly by signal, voting position, and treatment. 4. Vote order effect: difference in Yes-voting rates across positions 1, 2, and 3 among benevolent participants with white signals in the sequential protocol. 5. Learning effects: changes in individual voting behaviour across rounds, estimated via round fixed effects.

Cited by

No citation data available for this registration yet.

Keywords