PaperlayerView exact revision on arXiv

Source-linked AI summary

Rules Before Oracles: Auditable, User-Configurable Argument Selection for Deliberative Polling

Muntaser Syed, Markus Zanker, Marius Silaghi

arXiv:2608.23979v1cs.AIcs.CRcs.CYcs.GTcs.MA

TL;DR

In large deliberative polls, argument selection is often opaque, leaving voters unable to check the slate that shaped their exposure. The paper evaluates slates by reason coverage, how early coverage arrives, and captured endorsement. It reports that an unconstrained learned ranker’s entire completeness advantage over the rule is three points.

  • Problem

    Argument selection in large deliberative polls is usually neither public nor checkable, leaving voters unable to recompute or contest the slate shaping their exposure.

    Source: §1 · p. 1, §4.1 · p. 13

  • Method

    The paper models an alternative-based poll and evaluates slates by reason coverage, how early coverage arrives, and captured endorsement mass.

    Source: §3 · p. 7

  • Results

    Three points of completeness is the entire competitive advantage available to an unconstrained learned ranker over the rule.

    Source: §11 · p. 40

  • Takeaways & Limitations

    The relation-weight policy is a security control, making its setting a consequential choice rather than one for the system to decide.

    Source: §11 · pp. 39–40

    Because the electorate is simulated, reported quantities describe agents under the model and do not establish magnitudes in human electorates.

    Source: §12 · p. 40

Abstract

from arXiv · show

In a deliberative poll, once submissions outnumber what anyone will read, some mechanism chooses which arguments each voter sees, acquiring much of the decision; practice delegates it to opaque learned rankers, so a voter cannot recompute or contest the exposure that shaped their vote. We ask whether it can be a published rule over publicly recomputable evidence with parameters held by the voter, treating legibility as an admissibility condition on usable mechanisms, not an objective traded against accuracy. We formalise a poll over bipolar justification sets, judging a slate by reason coverage, the order it arrives in, and captured endorsement mass; we give seven checkable criteria for a civic recommender and a rule meeting them: a one-hop reversed endorsement flow parameterised by a relation-weight function. An agentic simulator records every slate at every vote, over about 17,000 seed-paired runs. Served slates fall 0.035 short of a label-reading ceiling upper-bounding every selection procedure, opaque ones included: any unconstrained ranker's advantage is bounded and small. On coverage alone, with non-degenerate authoring, the rule is indistinguishable from a random slate, a null due to an order-blind, charity-blind instrument; on the other two it leads at every prefix by a margin widening with adversarial pressure and dominates on mass by a factor of 3.3. Once a realistic fraction of submissions carries no reasons, the coverage margin returns and grows. Label-homogeneous flooding collapses completeness from 0.81 to 0.34 under a flat weight policy, only to 0.44 under author-count normalisation, making the weight function a security control worth 10% of completeness. The choice between ranking arms is a position on a coverage-versus-mass frontier, not a fact, the kind of choice only a legible rule can hand to the person it affects. It maps onto an open-source peer-to-peer platform.

1 Introduction

The paper argues that argument selection in civic polling must be a published, reproducible rule over public evidence with voter-held parameters, then evaluates a one-hop reversed-endorsement rule and its implementation.

1 Introduction

  • The proposed voter-configurable rule uses publicly recomputable evidence, while the paper defines seven checkable recommender criteria and measures coverage, order, and captured endorsement mass.The rule uses one-hop reversed endorsement flow over rebuttal and reinforcement links; its policy vector lets voters configure it without making it individually unpredictable.Source: §1 · p. 2

1.2 What the measurements turned out to say

  • 0.031 ± 0.014 coverage fraction distance separates served slates from the label-reading greedy ceiling, versus a 0.163 temporal component unrecoverable by any procedure.The ceiling upper-bounds every selection procedure on the same pool.Source: §1 · p. 2
  • Under non-degenerate authoring, coverage does not distinguish the rule from random selection; ordering favors the rule at every prefix, while degenerate submissions restore coverage gains.The coverage null is tied to an order-blind, charity-blind instrument; link summands are nearly inert under non-degenerate authoring but become valuable as unjustified submissions increase.Source: §1 · p. 2, §1.2 · p. 4

1.1 The exposure problem, and the position taken here

  • Large submission pools force voters to see selected slates, creating temporal inequity, silent narrowing, and influence that users cannot falsify under opaque learned selectors.The paper treats legibility as an admissibility condition for civic selection, not a performance objective.Source: §1.1 · p. 3

1.3 Why the substrate matters

  • The paper maps its design onto DDP2P (M. C. Silaghi, Alhamed, et al. 2013; M. C. Silaghi, Qin, Alhamed, et al. 2013), where local evaluation over peer-held replicas can make charter criteria structural facts.The platform uses public-key and content-digest identifiers and push–pull gossip to synchronize independent databases.Source: §1.3 · p. 4

2 Background and Related Work

This section positions per-voter argument selection as the scaling problem between deliberative polling and a voter’s ballot, then distinguishes the paper’s auditable, semantics-free selection approach from systems that judge arguments or rely on opaque serving-time inference. It also situates the work among recommendation, participatory decision-making, agentic simulation, and decentralized-platform research.

Deliberative polling and its scaling problem

  • Digital deliberative polls shift the task of maintaining balanced exposure from organisers to per-voter selection, which this work treats as an object to engineer.Prior studies report opinion movement after deliberation, especially among initially less-informed participants; digital deployments scale participation but cannot rely on everyone reading the full corpus.Source: §2 · p. 5

Bipolar argumentation

  • The paper uses bipolar attack-and-support links as evidence of relevance, but declines argumentation semantics that would decide which arguments a voter should disregard.This shallow use differs from dialectical-quality measures; the links structure relations rather than adjudicate truth (Dung 1995; Cayrol and Lagasquie-Schiex 2013, 2005; Gonzalez et al. 2021; Rocha et al. 2026; Prakken 2001; Robert 1915).Source: §2 · p. 5

Computational social choice for participatory processes

  • The paper addresses what voters see before forming positions, complementing computational social-choice rules that aggregate positions after participation.Those downstream rules can be paired with the slate mechanism.Source: §2 · pp. 5–6

Argument mining and its role here

  • Argument labels are extracted once at authoring time, attached to items, and open to contest and correction before serving; selection then uses that frozen public output.This places language models on the authoring side of the boundary, so disputes can target labels rather than model internals; adversarial labels are revisited in Section 12.Source: §2 · p. 6

Civic platforms with algorithmic assistance

  • Existing civic platforms surface bridge statements or infer participant values, but the paper argues that unauditable model reasoning is unsuitable for binding processes regardless of measured quality.It notes transparency and human-oversight duties in the Digital Services Act and AI Act, alongside concerns about sycophancy and political leaning (Small et al. 2021; Liscio et al. 2025; Brenneis et al. 2021; Mancini 2015; Behrendt et al. 2024; Tessler et al. 2024; European Parliament and Council 2022, 2024; Sharma et al. 2023; Feng et al. 2023).Source: §2 · p. 6

Diversity, coverage and manipulation in recommendation

  • A one-hop, non-iterative rule avoids the link-farm susceptibility of eigenvector-style ranking; the paper reports hub-riding is inert against it, while residual exposure remains.The rule’s completeness objective covers reasons live at vote time, extending recommendation coverage work to a changing civic context (Burke 2017).Source: §2 · pp. 6–7

Agentic simulation as an evaluation method

  • The agentic simulator supports claims about mechanisms under a stated behavioral model, not claims about the magnitude of effects in a human electorate.This caveat situates the evaluation alongside generative-agent and profile-conditioned survey-response research (Park et al. 2023; Argyle et al. 2023; Q. Wu et al. 2024; Bail 2024).Source: §2 · p. 7

Decentralised infrastructure for civic processes

  • The implementation builds on DDP2P, a gossip-replicated store of signed items for civic decisions, rather than proposing a new ledger or distributed hash table.Its surrounding research covers census verification, false-identity detection, reputation, recommendation, trust, and related infrastructure (Maymounkov and Mazières 2002; Demers et al. 1988; Shapiro et al. 2011; Merkle 1988; Nakamoto 2008; Nikitin et al. 2017; Miller 2020; M. C. Silaghi, Alhamed, et al. 2013; M. C. Silaghi, Qin, Alhamed, et al. 2013; M. C. Silaghi and Roussev 2014).Source: §2 · p. 7

3 The Object: An Alternative-Based Poll and Three Ways to Judge a Slate

The section defines an alternative-based poll and three complementary ways to judge a slate: reason coverage, how early coverage arrives, and captured endorsement mass. It also establishes a coverage ceiling and specifies what evidence the selection rule can use.

3.1 Polls, sides, and justifications

  • The poll models support and opposition justifications connected by typed rebuttal and reinforcement links, with public endorsement counts and configurable link weights.Choosing flat or distinct-author-normalised link weights changes completeness under coordinated flooding by 0.08–0.12.Source: §3.1 · pp. 7–8

3.2 Reason vocabularies and the live denominator

  • Completeness compares a slate with reasons already present when the voter votes, but cannot distinguish slate order or reward better arguments when all submissions are well-formed.Averaging the two sides can hide a poorly covered side, so the paper also evaluates completeness using the worse-covered side.Source: §3.2 · pp. 8–9

3.3 Three instruments: coverage, order, and endorsement mass

  • Slates are judged by live-reason coverage, how early coverage arrives, and the share of maximum endorsement weight captured.Coverage is order-blind and charity-blind; endorsement mass asks whether a slate shows what constituents adopted.Source: §3.3 · pp. 9–10

3.4 The Subsuming Justification Problem

  • The Subsuming Justification Problem is NP-complete; weighted and componentwise variants encode priority reasons and balanced coverage across both sides.The weighted variant can represent reasons that must not be crowded out, while the componentwise variant prevents one side from compensating for the other.Source: §3.4 · p. 10

3.5 The greedy ceiling, and the discipline governing its use

  • The offline greedy coverage ceiling is 0.031 ± 0.014 from served slates and bounds what any selector could achieve to within its 0.632 guarantee.It is an evaluation-only oracle: no agent saw its slates, and it is not a deployable selection rule.Source: §3.5 · p. 11

3.6 Reach: what a slate can be held responsible for

  • A peer can compute the rule’s scores from a partial replica using candidate items’ reach and outgoing links, but unendorsed cold-start items remain invisible.Those items surface through authoring-side random exploration, not by tuning the rule.Source: §3.6 · p. 11

3.7 A worked example

  • In the bus example, equal 0.60 coverage hides endorsement-mass differences: the rule captures 1.00, versus 0.64 for {s1,s5} and 0.19 for {s3,s5}.Replacing a reason-bearing item with an unendorsed, reasonless submission leaves coverage blind to the change, while the rule scores that submission zero.Source: §3.7 · p. 12

4 The Charter: Criteria Before Objectives

The section treats recomputability, contestability, and related procedural properties as conditions for using a civic selector, not qualities to trade against ranking performance. It sets out seven criteria and a simple endorsement rule that satisfies them.

4.1 What opacity costs, concretely

  • Opaque selectors obstruct independent recomputation, derivational attribution and explanation, terminating disputes, and reconstruction of historical slates.Under a published rule, disputes can resolve through public tallies, signed links, or open policy debate; rule changes are also recoverable as diffs.Source: §4.1 · p. 13

4.2 Seven criteria, with their tests

  • The seven checkable criteria include determinism, evidence locality, author blindness, and semantic abstinence at serving time.Their tests include rerunning a frozen snapshot, perturbing evidence outside the declared neighbourhood, permuting author identities, and replacing text with opaque identifiers.Source: §4.2 · pp. 13–14
  • The remaining criteria are reproducibility, contestability, and participant-held configurability, each paired with an implementation test.Tests include reconstructing a historical slate, decomposing its score into public-evidence contributions, and letting participants change their policy and observe the effect.Source: §4.2 · p. 14

4.3 Why this is not an empirical hypothesis

  • A learned selector’s ranking quality cannot compensate for failing the procedural requirements of a binding civic process.The section argues that these requirements are prior to measured quality; empirical comparisons instead assess the cost and consequences of choosing an admissible rule.Source: §4.3 · p. 15

4.4 The endorsement rule

  • The endorsement rule ranks items using one-hop reversed endorsement flow, crediting endorsements of the items they point to rather than incoming links.The rule uses public evidence and separately displayable score terms; labels evaluate slates but do not determine selection.Source: §4.4 · p. 15

4.5 The weight policy is the whole game

  • Changing the relation-weight policy shifts completeness by 0.08–0.12 under coordinated attack, making that policy a security control.Distinct-author normalisation limits the link credit a coalition can create by spreading its links; the passage reports that co-signing instead raises completeness monotonically by concentrating identities on fewer promoted items.Source: §4.5 · p. 16

4.6 Configurability without incoherence

  • Participants can choose different published policy parameters over identical public evidence, producing separately recomputable slates rather than opaque personalisation.Recording each participant’s policy with the slate preserves reproducibility and contestability; the coverage-versus-endorsement-mass choice is presented as a frontier, not a system-level optimum.Source: §4.6 · p. 17

4.7 Rules of order as algorithms of interaction

  • The section likens a published selection rule to rules of order: procedures for deliberation should be knowable in advance by those bound by them.It frames the screen’s published rule as a standing order for deliberation.Source: §4.7 · p. 17

5 Abas: An Agentic Instrument for Auditable Deliberation

Abas simulates deliberative polling and preserves enough evidence to reconstruct and audit every slate served. Its agent model, controlled comparisons, persistent record, and browser interface make the mechanism inspectable rather than treating the simulator as a product demonstration.

5.1 The constituent agent

  • Agents hold a scalar opinion that stochastically determines ballot direction and use TF–IDF cosine similarity between position text and existing material to judge what speaks for them.They do not update opinions in response to exposure, so results concern mechanism behavior under a stated model, not evidence about constituent behavior.Source: §5.1 · p. 18

5.2 The round protocol

  • Adoption is limited to served items, creating feedback through endorsement tallies, while independent authoring draws keep the corpus bit-identical across ranking arms.Agents choose rebuttal and reinforcement targets themselves, respectively selecting the least-similar opposing item and the most-similar same-side item.Source: §5.2 · p. 18

5.3 Submissions that fail to justify

  • Degenerate submissions are varied by share and error type, while dedicated randomness keeps the corpus skeleton bit-identical for paired comparisons.Reason-free items occupy slate slots without adding labels; erroneous supports add opposing-side labels, and completeness uses only canonical reasons for its denominator.Source: §5.3 · pp. 18–19

5.4 Ranking arms and the evaluation ceiling

  • The ranking arms differ only in the ranking step, and an evaluation-only greedy-cover ceiling reads labels directly, making it inadmissible as a deployable mechanism.The compared arms include the full rule, endorsement-only, enhance-only, attack-only, and random; random slates are seed-paired with the other arms.Source: §5.4 · pp. 19–20

5.5 Persistence: the record is the point

  • Every served slate is stored with its constituent, time, policy vector, and item order, enabling exact reconstruction and recomputation of scores against the ranking served.The database also records agents, justifications, links, ballots, parameters, and seeds, making ordering and endorsement-mass analyses computable without rerunning the experiment.Source: §5.5 · p. 20

5.6 Inspecting the record

  • A browser lets constituents inspect their served slates in order, score components, contributing links and authors, following ballot, and the counterfactual slate under another policy vector.The interface uses tables and links to present the recorded arithmetic without visualizing information beyond it.Source: §5.6 · p. 20

6 Experimental Setup

The section fixes the propositions, reference settings, adversary capabilities, and statistical conventions used to interpret the experiments; a third proposition is reserved for the refresh-interval scenario.

6.1 Propositions

  • UBI and BRA ask whether to address the AI revolution through basic income or basic resource assurance; OPT-OUT is used only for the refresh-interval test.The propositions use thirty canonical reasons per side; OPT-OUT concerns pupils studying without computers or internet access.Source: §6.1 · p. 21

6.2 Reference configuration

  • At the reference setting, N=1000 constituents use K=20 items per side, author-normalised weights, and score-cache refresh after every vote.The configuration also sets πown=0.10, πreuse=0.50, πabst=0.40, and πlnk=0.60; its corpus is roughly 50 items per side against 30 labels, with a near-even ballot split.Source: §6.2 · p. 21

6.4 Adversary model

  • Coalitions are spread through the processing order under authenticated voting, and attacks vary their authoring, linking, targeting, and label homogeneity.The model includes hub-riding, label flooding, heterogeneous controls, and co-signed flooding; the implementation permits repeated link assertions by distinct authors, leaving the co-signing numerator movable.Source: §6.4 · p. 22
  • Flooding uses a shared label pair, while the heterogeneous control independently draws label pairs and a targeting variant tests hub aiming.Co-signed flooding groups coalition members to raise the distinct-author numerator; group sizes C∈{1, 2, 5, 10, 25, 50} are tested at two coalition sizes under both weight policies, with C=1 reproducing the flooding sweep.Source: §6.4 · p. 22

6.5 Statistical conventions

  • Paired arms use two-sided paired t-tests, unpaired configurations use Welch’s test, and intervals report across-seed standard deviations unless noted.Raw p-values are reported for multiple contrasts; contrasts failing Bonferroni adjustment are called suggestive, manuscript-wide hypotheses are unadjusted, and effects near p≈10−2 warrant caution while core arguments rely on p<10−10.Source: §6.5 · p. 22
  • At N=1000, measured vocabulary saturation occurs at about 800 constituents for πown=0.02 and 150 for πown=0.30.The figure marks these measured anchors; the curve shape between them is illustrative, and its left-side area represents voters who voted while the reason space was still under construction.Source: Figure 2 · p. 23

7 Results I: An Attack-Free Electorate

In a non-degenerate, attack-free electorate, served slates cover about four-fifths of the reasons already written; most incompleteness comes from reasons not yet written, while the label-reading ceiling’s selection advantage is small.

7.1 What the reference configuration produces

  • 0.806 ± 0.038 on BRA and 0.804 ± 0.032 on UBI: reference slates cover about four fifths of the reasons written on each side.Within-run completeness spread is 0.213, more than five times the between-run standard deviation; arrival time, as the reason vocabulary assembles, is the dominant source of variation in what voters see.Source: §7.1 · p. 23

7.2 Which levers move coverage

  • Completeness spans 0.44 at K=5 to 0.83 at K=30; raising πown from 0.02 to 0.30 lifts it from 0.70 to 0.88, while link-rate changes shift it just 0.004.Electorate size also raises completeness, from 0.69 at N=200 to 0.90 at N=5000; authoring rate is the cheapest lever per point gained, while link rate mostly reshuffles arguments that would be served anyway.Source: §7.2 · p. 24

7.3 The price of semantic abstinence, measured

  • 0.806 ± 0.038 served completeness versus 0.837 ± 0.034 for the label-reading ceiling; live-vocabulary scores are 0.968 ± 0.015 versus 1.000.The 0.194 shortfall from complete coverage comprises a 0.031 selection component and a 0.163 temporal component; four fifths of measured incompleteness reflects sequential deliberation, not selection.Source: §7.3 · pp. 24–25, Table 5 · p. 25

8 Results II: What the Coverage Measure Could Not See

With non-degenerate authoring, union-based coverage can make ranking look no better than random, even as ranking improves what readers encounter early. When submissions include reason-free material, the rule gains on coverage; endorsement mass also reveals a trade-off between reason breadth and fidelity to endorsements.

8.1 The null result, for non-degenerate authoring

  • With non-degenerate authoring, completeness barely separates the rule and random slate without attack, while the random slate wins under coalition flooding.At zero attackers, completeness is 0.8205 for the rule against 0.8181 for random; at 0.1 and 0.25 electorate, the rule scores 0.6088 against random’s 0.6472 and 0.3731 against 0.4437. Under author normalisation, link bonuses never exceeded 0.96 against a K-boundary endorsement gap of 3, so rank changes could be invisible to union-only completeness.Source: §8.1 · p. 26, Table 6 · p. 26

8.2 Order: what a ranking rule is actually for

  • At every prefix before m=20, endorsement ranking leads the uniform draw, with gains widening under attack; at a quarter-electorate coalition, p5 rises by 0.1387 and AUC by 0.0796.Under that coalition, a reader reaches nine tenths of the ranked slate’s coverage after 8.9 items, compared with 14.6 for the uniform draw, although both end at the same p20.Source: §8.2 · pp. 26–27, Figure 3 · p. 27, Table 7 · p. 28

8.3 Degenerate authoring: giving the rule something to discriminate against

  • At G=0.7, the rule leads the uniform draw on combined completeness by 0.0471, with the margin growing from near-zero at G=0 as degenerate material increases.At G=0.7, the rule serves 66.6% degenerate material against the uniform draw’s 70.5% corpus share. Raising E shrinks the margin by 0.005 to 0.010 across the grid.Source: §8.3 · p. 28, Table 8 · p. 29

8.4 Why the link terms square the discrimination

  • At G=0.7, genuine items lead reason-free items by a factor of 7.7 on endorsements and 7.2 on link credit, so link terms square existing discrimination.This mechanism depends on degenerate items receiving fewer adoptions in the simulator: 1.43 endorsements against 11.04, driven by agents’ TF–IDF similarity-based adoption. The finding therefore concerns how the rule inherits and amplifies electorate discrimination, whose rate among human voters was not measured.Source: §8.4 · p. 29

8.5 Coverage against endorsement mass

  • At a quarter-electorate coalition, the uniform draw captures 0.17 of achievable endorsement mass against the rule’s 0.55, a factor of 3.3, despite leading on coverage.The greedy ceiling reaches coverage 1.0000 in every regime, but under a quarter-electorate coalition it has endorsement mass 0.38 against the rule’s 0.55. At G=0.5, the link terms buy +2.8 points of coverage for −2.4 points of endorsement mass, a frontier rather than a single best setting.Source: Table 9 · p. 30, §8.5 · p. 30

9 Results III: A Coalition in the Electorate

With authentication, a coalition can manipulate what it writes and links but cannot vote twice or forge endorsements. The results distinguish harmless hub-riding from damaging label-homogeneous flooding, which author normalisation mitigates but does not eliminate.

9.1 Hub-riding is inert

  • Hub-riding leaves completeness statistically unchanged at every coalition size on both propositions (all p > 0.50), with the two weight policies differing by less than 0.002.The one-hop rule gives a bounded bonus without compounding, so the attacker’s argument is seen while slate informational quality remains intact.Source: §9.1 · pp. 30–31

9.2 Label flooding, and the weight policy as a security control

  • At 25% coalition size, completeness falls from 0.816 to 0.377 on BRA and 0.815 to 0.372 on UBI under author normalisation, versus 0.228 and 0.219 under flat weights.The author-normalisation defence margin is 0.117 at 5% coalition size, widens to 0.181 at 15%, then eases to 0.151 at 25%.Source: §9.2 · p. 31, Table 10 · p. 32

9.3 The mean over sides hides how one-sided the damage is

  • At 15% coalition size, mean completeness of 0.52 coexists with 58% of constituents having a side below 0.5 under author normalisation, and 94% under flat weights.The worse-side measure registers 5–13% more damage than the mean across coalition sizes; where a deployment sets a certification threshold, the paper argues it belongs on imbalance, not the mean.Source: §9.3 · pp. 31–32, Table 11 · p. 32

9.4 The refresh interval matters only when someone is attacking

  • At M = 200, lagging refresh raises the ranked arm’s completeness by 0.013 under attack and leaves the uniform draw unchanged; without a coalition, the interval changes nothing the ordering claims rest on.The paper reports at M = 1 as the conservative choice: the shorter interval lets the flood into ranked slates slightly faster, while the defence remains counting distinct authors.Source: Table 12 · p. 33, §9.4 · pp. 33–34

9.5 The collapse is homogeneity, not hyperactivity

  • With rate- and corpus-matched controls, label homogeneity accounts for 78%–85% of the completeness collapse, while volume damage alone costs 0.067 at a quarter-electorate coalition under normalisation.Hub-aiming is negligible under normalisation (≤0.006 at every coalition size, all p ≥ 0.14); the pipeline cannot absorb many voices saying one thing.Source: §9.5 · pp. 34–35, Table 14 · p. 35

9.6 The coalition cannot coordinate its way out

  • Completeness rises monotonically with coordination group size in every tested cell; at 25% coalition size, a fully coordinated coalition leaves completeness at 0.759 against a 0.815 baseline.On BRA at 25%, increasing group size from 1 to 50 multiplies mean co-signers per edge by 3.4 but cuts poison items from 250.1 to 10.0; the gap between weight policies narrows from 0.149 to 0.005.Source: §9.6 · pp. 35–36, Table 15 · p. 36

9.7 What the adversarial results say about the charter

  • The defence is a published distinct-author policy, not a detector, and co-signing makes the coalition’s coordination visible in the public link record.The section locates residual risk in corpus capture: homogeneity causes four fifths of the damage, so it points to census integrity and slate capacity rather than ranking as mitigations.Source: §9.7 · p. 36

10 Peer-to-Peer Realisation on DDP2P

DDP2P supports peer-local evaluation of a configurable, auditable slate rule, while making incomplete evidence and identity verification explicit. In this setting, ordering, link contributions, and voter-selected weight policies become especially consequential.

What the platform provides

  • DDP2P provides self-contained, globally identified items and push–pull gossip without a global index or authoritative replica.Directory and data servers are replaceable and their contents verifiable; organisations define constituencies and jurisdictions through rules rather than accounts.Source: §10 · p. 37

Mapping the model onto the platform

  • The model maps closely to DDP2P, but labels λ and slates S_i are new: authors sign labels, and slates can remain local or be published as signed receipts.Labels are claims that readers can check against text and dispute; mining systems may propose them, but proposals are advisory rather than ground truth.Source: §10 · p. 37, Table 16 · p. 38

Identity

  • Peer-to-peer membership makes identity verification a system problem: a witnessed, signed census estimates the real eligible population, while observers choose their own criteria over shared evidence.Under authenticated identities, coalitions cannot inflate E, leaving the link graph as the remaining attack surface; census integrity therefore constrains coalition size.Source: §10 · pp. 37–38

Evaluating the rule on an incomplete replica

  • On an incomplete replica, missing items can reduce completeness but cannot corrupt held-item scores, which remain lower bounds when votes or links are missing.Prioritising votes and links, reporting replica coverage, and append-only union merges make incompleteness visible as replicas converge; the peer trades complete evidence for verifiable local computation.Source: §10 · pp. 38–39

Implications for a peer deployment

  • In peer deployments, first-five ordering and link contributions matter more, while a voter-local weight policy is a security control worth ten points of completeness.Each participant can choose ωr and compare alternatives; the coverage-versus-mass frontier has no uniquely correct choice.Source: §10 · p. 39

11 Discussion

The discussion argues that legibility has a bounded, small cost, while instrument choice and published policy shape what evaluation and robustness show. It treats ranking tradeoffs as choices for participants, not neutral decisions for the platform.

11 Discussion

  • Semantic abstinence costs 0.031 ± 0.014 against the all-procedure ceiling; an unconstrained ranker’s three-point completeness gain is smaller than 0.10 from higher authoring or 0.14 from doubling slate size.Four fifths of observed incompleteness was vocabulary that did not yet exist, which no procedure could recover. The cited comparison gives the authoring-rate change as 0.02 to 0.10.Source: §11 · pp. 39–40
  • Across 4000 seed-paired runs, a set-functional evaluation made the rule indistinguishable from a random subset, showing that coverage-only aggregates can hide selector differences.Remarks 1–2 predicted this measurement artefact; the discussion notes that coverage-style aggregates are common in diversity-aware recommender evaluations.Source: §11 · p. 40
  • Author-count normalisation is the one-line defence: it is worth 0.151–0.181 of completeness against a coalition holding a fifth of the electorate, without scoring trust or removing material.It counts distinct authors per side, making coordination visible in the public link record while preserving author blindness; authentication relocates residual risk to corpus capture, census integrity, and slate capacity.Source: §11 · p. 40
  • At G= 0.5, link summands buy 2.8 points of coverage for 2.4 points of endorsement mass, a frontier tradeoff rather than a uniquely correct setting.The discussion presents participant-held parameters as necessary because choosing between breadth of represented reasons and fidelity to constituent uptake is a political choice.Source: §11 · p. 40
  • Deployment should prioritize authoring and publish the weight policy; participant-held parameters also align with existing disclosure and oversight duties (European Parliament and Council 2022; European Parliament and Council 2024).The discussion says selection shortfall is a fifth the size of the temporal shortfall and the weight policy has an order-of-magnitude effect under attack; the cited laws require recommender-parameter disclosure and a non-profiling option, and transparency and human oversight for democratic-process systems.Source: §11 · p. 40

12 Limitations

The conclusions are conditional on a simulated electorate and model-specific adoption behavior; they do not establish human-electorate magnitudes. Key deployment assumptions and fairness questions remain untested.

12 Limitations

  • Every reported number describes simulated agents, not people, and the model omits opinion change, boredom, persuasion, and strategic behavior beyond coalitions.Adoption is TF–IDF similarity to position text, so the results support only mechanism claims under the stated model.Source: §12 · p. 40
  • Reason-free submissions receive 1.45 endorsements against 10.94 for reasoned items, so the coverage margin depends on the simulator’s adoption behavior.If people endorsed reason-free and reasoned submissions equally, link summands would become inert and the coverage margin would disappear; measuring human adoption is the authors’ priority follow-up.Source: §12 · p. 41
  • The ordering advantage remains untested under opinion-revising multi-round deliberation and genuinely divergent peer replicas.The multi-round implementation was not thoroughly evaluated, and no experiment used peers with partial replicas.Source: §12 · p. 41
  • The manuscript reports no manuscript-wide multiplicity correction; the two link-term effects at p≈10−2 are suggestive, while effects relied on are at p< 10−10.The authors state that the latter effects would survive any reasonable correction.Source: §12 · p. 41
  • Sybil resistance is assumed rather than demonstrated, and a broken census invalidates every Section 9 result.The proposed DDP2P witnessed-census approach was not evaluated (Qin, M. C. Silaghi, Hussien, et al. 2014; Qin, M. C. Silaghi, Matsui, et al. 2013b; M. C. Silaghi, Qin, Matsui, et al. 2016).Source: §12 · p. 41
  • Whether endorsement mass systematically disadvantages minority positions is an open question that the weighted variant does not answer.Minority reasons may appear in fewer items and therefore accumulate less endorsement.Source: §12 · p. 41

13 Work, Meaning, and the E-Citizen

The section argues that procedure makes collective agency possible, while AI’s release of productive labour could make substantial citizenship available to many—provided civic instruments remain legible to those using them.

13 Work, Meaning, and the E-Citizen

  • Procedural rules make collective reasoning possible by constraining form rather than content, and digital selection now requires that same legibility, configurability, and contestability in software.The section traces these principles to rules of order, authority over assembly rules, and appeals from the chair.Source: §13 · pp. 41–42
  • Athenian self-government depended on citizens’ time, while representative participation economises on it and can become a thin, performative gesture.The section links this history to opaque argument exposure as a kitsch appearance of meeting arguments without actually encountering them (Narcisse 2012; Bright and Margetts 2016; Toots 2019; M. C. Silaghi 2025; Calinescu 1987; Lazare 1999).Source: §13 · p. 42
  • One-in-ten reason-writing rather than one-in-fifty was worth as much completeness as multiplying the electorate fivefold, making public reasoning hours—not voter count—the binding constraint.The section argues that inclusive deliberation depends on cognitive diversity actually reasoning, and therefore measures reason-vocabulary coverage rather than agreement (Landemore 2013, 2012).Source: §13 · p. 42
  • 0.031 was the entire measurable completeness cost of preserving recomputable exposure, a trade-off the section rejects in favour of criteria C4 and C7.Without recomputability, citizens may gain leisure while becoming spectators to technology that assembles arguments and filters objections.Source: §13 · p. 42
  • Language models may help authors express their intended reasons, but must not unaccountably choose which reasons voters encounter at serving time.The section places model-assisted extraction at authoring time, where it is visible to and contestable by the author.Source: §13 · p. 43

14 Conclusions

The paper defends a published, voter-configurable selection rule over public evidence as democratic procedure: its one-hop reversed endorsement flow satisfies seven admissibility criteria. It argues that voters should be able to recompute why they were shown an argument from a table of numbers, at a measured cost of about three points of coverage.

14 Conclusions

  • The construction is an alternative-based poll over bipolar justification sets, served by a one-hop reversed endorsement flow with a relation-weight function as its only policy lever.The rule satisfies all seven stated admissibility criteria; the conclusion’s central defense is that exposure should be explainable through recomputable numbers, not narrative.Source: §14 · p. 43
  • The semantic-abstinence price was 0.031 ± 0.014 against a ceiling bounding every mechanism, with four fifths of residual shortfall temporal rather than algorithmic.The measurements span roughly 17,000 seeded runs.Source: §14 · p. 43
  • On non-degenerate authoring, set coverage cannot distinguish the rule from a random draw; order-sensitive performance leads at every prefix short of the full slate, reaching −8.3 positions of 𝑒90 at a quarter-electorate coalition.The order-sensitive margin widens under attack.Source: §14 · p. 43
  • When a realistic fraction of submissions fails to justify, the coverage margin returns and grows monotonically; link summands supply 93% of it at 𝐺= 0.7.The link summands do so by squaring the electorate’s own discrimination.Source: §14 · p. 43
  • On endorsement mass, the random baseline loses by a factor of 3.3, while the greedy ceiling stops dominating under attack.Against label flooding, the weight policy is worth 0.151–0.181; four fifths of the damage is homogeneity rather than volume.Source: §14 · p. 43
  • Hub-riding is inert, while a co-signing coalition makes itself monotonically weaker.Label flooding remains severe despite the weight policy’s benefit.Source: §14 · p. 43
  • The mechanism’s results depend on whether human voters discriminate against unjustified submissions at the modelled rate and on census integrity for every adversarial result.Neither assumption was established.Source: §14 · p. 43

A note on scope and provenance

This manuscript is an original, independently written extended treatment, not a camera-ready conference paper or a reproduction of one. Its overlap with prior work is limited to subject matter and numerical results from the same runs; it develops several extensions.

A note on scope and provenance

  • Every definition, proposition, algorithm, example, figure, and table in the manuscript was composed for this document.The manuscript also introduces its own notation, three-instrument framing, seven criteria and tests, Section 3.7 worked example, and diagrams.Source: A note on scope and provenance · p. 43
  • Overlap with prior reports is limited to subject matter and numerical results computed from the same experimental runs.Results attributed to prior work are cited as such.Source: A note on scope and provenance · p. 43
  • Extensions include ordering and endorsement-mass instruments, degenerate authoring and its mechanism analysis, and the coverage-versus-mass frontier.The manuscript also elaborates the criteria and audit tests and evaluates side balance.Source: A note on scope and provenance · p. 43

Tools, data and reproducibility

The simulations are reproducible from recorded seeds, parameters, served slates, and policy vectors, with run databases and analysis scripts publicly available. The authors used language-model assistance only for drafting and editing prose.

Tools, data and reproducibility

  • Each seeded simulation records its seed, full parameter set, served slates in order, and policy vector; run databases and analysis scripts are publicly available.The materials are hosted at https://github.com/devfitcs/ABAS.Source: Tools, data and reproducibility · p. 44
  • Language-model assistance was limited to drafting and editing prose; technical content, experimental design, analysis, and conclusions are the author’s.Source: Tools, data and reproducibility · p. 44

A The Comparison This Manuscript Declines to Run · A.1 What would make such an experiment worth running

The appendix rejects a learned-ranker benchmark because the paper’s claim concerns civic admissibility, not comparative performance, and proposes a human-participant experiment to test whether legibility does procedural work.

A The Comparison This Manuscript Declines to Run

  • The appendix records what a learned-ranker experiment would need to test, rather than presenting a performance comparison the paper would not act on.The manuscript argues that procedural properties can take priority over measurement quality, as with the secret ballot, bookkeeping rules, and rules of order.Source: A The Comparison This Manuscript… · pp. 47–48
  • A learned ranker’s coverage, ordering, or engagement advantage cannot establish admissibility, because even superior coverage would leave key civic criteria unmet.The claim is normative: a mechanism that cannot be recomputed, attributed, contested, or reconstructed is unsuitable for a binding civic process.Source: A The Comparison This Manuscript… · pp. 47–48
  • 0.035 ± 0.013 bounds the entire competitive advantage available to an unconstrained selection mechanism against a label-reading greedy cover.The ceiling saturates the live vocabulary in every seed and upper-bounds what any selection procedure, including a learned ranker, could achieve on the same pool.Source: A The Comparison This Manuscript… · p. 48
  • A learned-ranker comparison is uninformative because its behaviour depends on opaque training data, objectives, and tuning choices.A result favouring either the proposed rule or the learned ranker would be unconvincing, and acting on the result would treat legibility as a tradeable quantity.Source: A The Comparison This Manuscript… · p. 48

A.1 What would make such an experiment worth running

  • A matched-population experiment could test legibility by giving groups the same corpus and served items, varying only whether selection is recomputable and decomposable.It would measure slate contests, whether contests terminate, and reported trust in the outcome; results could change the authors’ position or its strength.Source: §A.1 · p. 48
  • The proposed human-participant test is out of scope for the simulation study and is identified as the most valuable experiment for future work.The authors place it on the roadmap alongside field deployment.Source: §A.1 · p. 48

Related papers

Loading 2608.23979v1…