Endorsement Without New Evidence: How Sequential Voting Inflates Mandates in Online Community Governance

Zihan Chen; Lei Nico Zheng; Di Zhu

Source-linked AI summary

Endorsement Without New Evidence: How Sequential Voting Inflates Mandates in Online Community Governance

Zihan Chen, Lei Nico Zheng, Di Zhu

arXiv:2609.09321v1Sep 8, 2026cs.SIcs.CYcs.HC

TL;DR

Large support margins in public community elections are often treated as strong mandates, although late votes may add less independent scrutiny to the record. The paper measures vote–text divergence in Wikipedia admin-election rationales and tests how it varies with arrival position and prior discussion. Divergence rises among later voters, is consistent with information saturation, and does not predict worse observable post-promotion outcomes.

Problem
Large support margins are read as stronger consensus and mandates, but late voters can add votes without adding independent scrutiny to the record.
Source: §1 · p. 1
Method
The paper measures vote–text divergence from rationale substance using five text dimensions and defines it as the negative residual after accounting for vote sign, decisiveness, and log length.
Source: §3.2 · p. 4
Results
Divergence rises with arrival position, tracks accumulated prior text rather than tally decisiveness, concentrates among network-peripheral voters, and does not predict worse observable post-promotion outcomes.
Source: §1 · p. 2
Takeaways & Limitations
Margins can overstate the scrutiny they signal even when selection appears unharmed on observable post-promotion outcomes.
Source: §1 · p. 2
The outcome evidence covers observable post-promotion activity and survival and cannot rule out unobserved-quality channels; the design also does not establish a clean causal effect of seeing the tally.
Source: §6.1 · p. 9

Abstract

from arXiv · show

Online communities often treat large support margins in public elections as strong mandates. We argue that such margins can overstate the independent scrutiny behind a decision. Using 198,275 free-text rationales from Wikipedia admin elections, we introduce vote-text divergence, a measure that flags a decisive vote paired with a thin, deferential rationale. Divergence rises as voters arrive later, even after controlling for voter and election fixed effects. The pattern is consistent with information saturation: once prior text is accounted for, arrival order no longer predicts divergence, while accumulated prior evidence does. The effect is strongest among peripheral voters in the co-voting network. Yet divergence does not predict worse post-promotion outcomes, such as administrative activity or survival. Public tallies can therefore weaken the scrutiny signal even while selecting capable administrators: a margin may appear to reflect more consensus and support than it actually contains.

1 Introduction

Public support margins are commonly read as strong mandates, but they may count votes without counting the independent scrutiny behind them. The paper measures this gap in Wikipedia admin elections and finds that later votes are more deferential, consistent with information saturation.

1 Introduction

Large support margins are often treated as evidence of community consensus and a stronger mandate, yet late voters may add votes without adding independent scrutiny.The margin may also inform bureaucrats’ judgments of legitimacy and later members’ assessments of a candidate’s reputation [6] [7] [8].Source: §1 · p. 1
The study introduces vote–text divergence to flag decisive votes paired with thin, deferential rationales, using written reasons to observe when vote direction and evidentiary contribution diverge.Unlike measures that infer conformity from actions or ratings, this approach uses the reasons recorded beside votes.Source: §1 · pp. 1–2
The analysis uses signed Wikipedia Requests for Adminship votes and links elected administrators to their later public administrative actions.The corpus contains 198,275 rationales across 4,003 elections from 2003 to 2013.Source: §1 · p. 2
Vote–text divergence rises with arrival position, tracking accumulated prior text rather than tally decisiveness, and concentrates among network-peripheral voters.Once prior text is controlled, arrival position no longer predicts divergence; the pattern is consistent with information saturation.Source: §1 · p. 2
Election-level divergence is nearly orthogonal to winning margins and does not predict worse observable post-promotion outcomes, while selection of capable administrators holds up.The authors report no deterioration from activity volume to removal for cause; the evidence concerns outcomes they can measure.Source: §1 · p. 2

2 Related Work and Theory

The paper distinguishes visible-vote influence from the loss of independent evidence as discussion accumulates, then develops a framework for measuring and testing that distinction. Its hypotheses predict a later-vote divergence gradient, concentrated among peripheral voters, without assuming downstream harm.

2.1 Governance by visible voting in online communities

The paper shifts attention from who wins online governance votes to whether larger margins reflect more independent scrutiny or more deferential repetition.Prior work considers promotion, deliberation, and sequential social networks in Wikipedia’s Requests for Adminship [6] [7] [8].Source: §2.1 · p. 2

2.2 Observational learning, saturation, and pivotality

Bandwagon influence follows a visible running count and reduced pivotality, whereas information saturation follows costly processing of an expanding discussion and perceived coverage of relevant evidence.Both can lower the incentive to add scrutiny, but saturation concerns the accumulated discussion record.Source: §2.2 · p. 3
Even efficient deference can concentrate independent scrutiny in early votes, so margins cease to count independent evidentiary contributions one for one.The measure captures the visible deficit but cannot distinguish efficient repetition avoidance from empty deference.Source: §2.2 · p. 3

2.3 Who defers: experience and motivation

The framework predicts that deference falls disproportionately to peripheral voters, who may have less incentive to supply costly scrutiny than invested core members.The proposed connection draws on findings about evaluator experience and contribution motives.Source: §2.3 · p. 3

2.4 Hidden profiles and text as data

Short conventional rationales create a validity challenge because wording and length may reflect convention or politeness rather than the substance of a voter’s contribution.The paper treats this as a construct-validity problem and tests the measure against alternatives.Source: §2.4 · p. 3

2.5 Hypotheses

The hypotheses predict that divergence rises with arrival position, is mediated by cumulative prior discussion, and is concentrated among low-centrality voters.The paper also tests whether election-level divergence predicts weaker downstream outcomes.Source: §2.5 · pp. 3–4

3 Data and Measurement

The study uses Wikipedia admin-election rationales to measure the substance accompanying each vote and links elected administrators to observable later outcomes. Its divergence measure separates deference and evidentiary substance from brevity alone.

3.1 The wiki-RfA corpus

The wiki-RfA corpus contains 198,275 votes across 4,003 elections, with rationale text for 96.4% of votes and timestamps parsed for 91.7%.The median rationale is 14 tokens, motivating a measure focused on effort and substance rather than long-form sentiment.Source: §3.1 · p. 4

3.2 Vote–text divergence

A transparent rule-and-lexicon system scores rationales for valence, certainty, specificity, grounding, and compliance with deference conventions.Informative conviction combines standardized valence, certainty, and specificity; divergence is its negative residual after accounting for vote sign, decisiveness, and log length.Source: §3.2 · p. 4
Vote–text divergence distinguishes a decisive, deferential, evidence-free vote from specific scrutiny, without simply relabeling short comments as divergent.The construct includes ambiguous cases, such as a short vote citing a prior argument or a generic claim of review; the authors identify human-coded validation as an important next step.Source: §3.2 · pp. 4–5

3.3 Observable governance outcomes

For elected administrators, the study builds an activity-and-survival index from public logs, combining standardized activity, breadth, tenure, sustained activity, and removal status.These observable outcomes are not a validated measure of administrator quality.Source: §3.3 · p. 5

4 Empirical Strategy

The empirical strategy tests whether arrival position predicts divergence within elections and whether prior tallies or accumulated text better account for the pattern. It also tests whether the gradient varies with network centrality and whether divergence relates to governance outcomes.

4 Empirical Strategy

Moving from the first to last voter is tested using election, voter, and two-way voter-and-election fixed effects, with within-election permutation and alternative-composite checks.The two-way specification absorbs stable voter writing tendencies and election-level conditions.Source: §4 · p. 5
To distinguish bandwagon from saturation accounts, the models add the support share and threshold cushion voters see, then cumulative prior text split into evidence and deference.This compares tally state with the accumulated discussion and its components.Source: §4 · p. 5
The peripheral-voter test interacts arrival position with co-voting-network eigenvector centrality and repeats the analysis using lagged election participation as an embeddedness proxy.The lagged measure checks against look-ahead in the static centrality measure.Source: §4 · p. 5
Governance-outcome models relate election-mean divergence to outcomes while controlling for margin, size, and year; additional tests use a promotion-threshold discontinuity and bureaucrat discretion.The fuzzy regression discontinuity instruments election with crossing the discretionary 75% promotion threshold.Source: §4 · pp. 5–6
Figure 1 describes a within-election pattern: later votes carry more deference language and cite less evidence.It presents the gradient’s two textual components.Source: Figure 1 · p. 6

5 Results

Later votes are more deferential and less evidence-grounded, and accumulated prior evidence explains the pattern better than visible tally decisiveness. Divergence is strongest among peripheral voters, yet does not predict worse observable governance outcomes.

5.1 Deference rises with arrival order (H1a)

Moving from the first to last voter raises divergence by 0.33 under election fixed effects and 0.45 under two-way voter-and-election fixed effects.Later arrival also raises deference language (+0.119) and lowers evidentiary grounding (−0.080) and specificity (−0.061), all with p < 0.001.Source: §5.1 · p. 6

5.2 Information saturation, not consensus (H1b)

The gradient does not strengthen with a more decisive tally: it is steeper in contested races (0.45) than landslides (0.25).A larger cushion above the promotion threshold predicts less divergence (−1.33, p < 0.001), weighing against visible-majority explanations.Source: §5.2 · p. 6
Controlling for prior discussion reduces the arrival-position effect from 0.33 to +0.02 (p = 0.54), while prior text predicts divergence (+0.09, p < 0.001).Prior evidence predicts later divergence (+0.34, p < 0.001), whereas prior deference language does not; the pattern is consistent with information saturation.Source: §5.2 · pp. 6–7

5.3 Deference is a peripheral-voter act (H2)

Arrival position predicts a steeper divergence gradient for network-peripheral voters than for central voters, with a position-by-centrality coefficient of −0.16 (p < 0.001).The interaction remains negative using strictly lagged embeddedness (−0.08, p < 0.001); peripheral status is a proxy for weaker scrutiny incentives, not direct proof of non-pivotality.Source: §5.3 · p. 7

5.4 No detectable deterioration in observable post-promotion outcomes

Election-mean divergence does not predict the activity-and-survival index (−0.011, p = 0.51), and the governance index shows no jump at the 75% threshold.The fuzzy-RD first stage is sharp, with pass rates rising from 17.8% to 81.1%; the instrumented governance-index effect is statistically zero across bandwidths.Source: §5.4 · p. 7, Table 4 · p. 7, Figure 2 · p. 8
Bureaucrats do not discount divergence when exercising discretion in grey-zone cases; estimates are near zero and powered to detect about ten-percentage-point changes per standard deviation.The divergence coefficient on promotion is small and insignificant across bands and under linear and logistic models.Source: §5.4 · pp. 7–8, Table 5 · p. 8

5.5 Alternative explanations and construct validity

Text alone cannot fully separate efficient deference to evidence already on the record from conformity to accumulated evidence, leaving construct validity unresolved.The paper identifies human coding as the remaining tool for making that distinction.Source: §5.5 · p. 8

5.6 Robustness

Among support votes in races ending above 90% support, divergence still rises with arrival position (+0.31, p < 0.001), and the consequences null survives robustness checks.Logistic models agree (all p > 0.36); the minimum detectable effect at 80% power is about 0.10 promotion probability per SD of divergence.Source: §5.6 · p. 8

6 Discussion

Public voting margins can overstate the independent scrutiny behind a decision, even when observable post-promotion outcomes do not worsen. The discussion argues for separating new evidence from endorsement and preserving independent evaluation before aggregation.

6.1 Theoretical implications

The margin can overstate independent scrutiny without detectable deterioration in observable outcomes, because deference is concentrated among peripheral voters with weaker incentives to scrutinize.The authors bound the outcome claim to measured post-promotion behavior and note that unobserved-quality channels remain possible.Source: §6.1 · p. 9
The authors do not claim a clean causal effect of seeing the tally; their evidence is consistent with deference produced by sequential public aggregation.Their outcome measures capture observable activity and survival, not unobserved quality.Source: §6.1 · p. 9

6.2 Design implications

Platforms should distinguish new evidence from endorsement, summarize distinct claims, and display evidence contributors alongside the raw support share.Collapsing repeated endorsements may prevent agreement from mimicking independent scrutiny; hiding the running count is secondary because the discussion record still accumulates.Source: §6.2 · p. 9
Independent rationales should be committed before agreements and disagreements are surfaced, since tally visibility alone does not address saturation of the discussion record.The authors describe hiding or coarsening the count as secondary to preserving independence before aggregation.Source: §6.2 · p. 9
Automated summaries may deepen saturation by making it easier for later voters to align with an existing consensus, while pre-digested context can enter the aggregate unchallenged.The paper’s pre-AI baseline is useful for measuring the human dynamic before that amplification.Source: §6.2 · pp. 9–10

7 Conclusion

The conclusion argues that sequential public voting can make a margin look more evidentially grounded than it is. The scrutiny signal weakens even as the resulting choice holds up on observable outcomes.

7 Conclusion

Later votes tend to defer rather than report new evidence, so the margin counts votes without counting the scrutiny behind them.The pattern tracks accumulated public discussion rather than tally decisiveness.Source: §7 · p. 10
The pattern holds under demanding voter and election fixed effects and is concentrated among network-peripheral voters.The conclusion says divergence tracks accumulated discussion rather than tally decisiveness.Source: §7 · p. 10
The margin is weakened as a measure of independent scrutiny, while the resulting choice holds up on the outcomes the study can observe.The conclusion connects this distinction to how platforms display and aggregate sequential records.Source: §7 · p. 10