I’m going to treat direct democracy as a chain—who sets the agenda, who develops the proposal, who decides, and who controls implementation—because most systems called “direct democracy” give citizens only the final vote. I’ll compare the strongest real cases with failed or cautionary ones, then assess the mob-rule claim against the evidence rather than treating it as either obvious or imaginary.

A pattern is already becoming clear: the best results do not come from maximizing the number of yes/no votes. They come from separating broad participation from the intensive work of learning, drafting, testing tradeoffs, and revising—then reconnecting those stages through transparent rules and a real government commitment to respond. I’m checking where that chain held and where governments broke it.

Direct democracy: what the evidence actually suggests

The evidence supports neither the utopian story—“let the people vote on everything and politics will become wise”—nor the nightmare story—“majority voting inevitably becomes mob rule.”

The strongest conclusion is:

Direct democracy works best as a constitutional pipeline for agenda-setting, deliberation, decision, and accountability—not as a succession of context-free yes/no polls.

Binding votes can correct elite drift and sometimes improve policy alignment. They can also produce discriminatory outcomes, badly drafted law, voter fatigue, and government-by-campaign-budget. Deliberative and participatory mechanisms improve the process, but only if their recommendations have a credible route into law.

1. “Direct democracy” covers several very different systems

A country can be highly direct at one stage and completely closed at another.

Most systems called direct democracy democratize only authorization. Officials, campaign organizations, or wealthy sponsors still decide what gets proposed and how it is worded.

That distinction is fundamental:

• A referendum lets the public accept or reject something developed elsewhere.
• A popular initiative lets citizens place a proposal on the ballot, although professional organizations often do the actual drafting and signature gathering.
• A citizens’ assembly gives a representative sample time to learn and develop recommendations, but usually lacks binding power.
• Participatory budgeting gives residents authority over a defined part of public spending.
• A town meeting can combine deliberation and binding decision-making, but generally only at local scale.

International IDEA’s comparative handbook is a useful global survey of referendums, initiatives, agenda initiatives, recall, campaign regulation, and institutional variation across more than 200 jurisdictions: Direct Democracy: The International IDEA Handbook.

2. Where it has been tried

Switzerland: the most developed national system

Switzerland combines:

• Mandatory referendums for constitutional amendments.
• Optional referendums: 50,000 signatures within 100 days can force a vote on legislation.
• Popular initiatives: 100,000 signatures within 18 months can propose a constitutional amendment.
• Extensive cantonal and municipal direct democracy.

About 200 optional federal referendums have been held since the mechanism was introduced in 1874; roughly 40% resulted in rejection of the challenged law. Constitutional changes require both a majority of voters and a majority of cantons. Swiss government explanation

Its most consequential effect may be indirect. Because legislation can be challenged, the government and Parliament negotiate with organized groups before enactment. Scholars sometimes call this the “referendum threat”: even when no vote occurs, the possibility of one changes drafting.

But Switzerland also demonstrates the limitations:

• Citizens usually vote on proposals rather than jointly developing them.
• Frequent voting rewards politically attentive, higher-resource participants.
• Wealthy organizations have advantages in signature gathering and campaigning.
• Some votes concerning immigrants and religious minorities have produced discriminatory outcomes.
• Federal double-majority rules give residents of small cantons more effective voting power.

Claims that direct democracy simply shrinks government are also too broad. Research controlling for voter preferences found approximately 8% lower cantonal expenditure but 20% higher local expenditure, with no clear change in total combined expenditure. Direct democracy appeared partly to decentralize spending rather than simply reduce it. Funk and Gathmann

United States: state initiatives and referendums

About half of US states have some form of citizen initiative or referendum, with California the best-known case.

This has produced genuine popular interventions where legislatures would not act. It has also exposed recurring problems:

• Paid signature-gathering means formal citizen access may function as market access.
• Ballot language frequently conceals substantial legal complexity.
• Voters can mandate spending while separately restricting taxation.
• Constitutional initiatives can lock detailed policy into rules that elected legislators cannot easily repair.
• Multiple overlapping initiatives may be internally inconsistent.
• Low-information elections increase dependence on party cues, advertising, and endorsements.

Campaign money does not mechanically purchase every result, but it strongly influences which proposals qualify, which frames dominate, and whether obscure measures can reach voters at all.

New England town meetings

Traditional town meetings in Vermont and other New England states place local legislative decisions before assembled residents. They come close to the classical ideal: participants can amend proposals, question officials, deliberate, and vote.

They work best when:

• The population is small.
• Participants share a concrete jurisdiction.
• Decisions concern visible local services.
• Meetings recur, so reputations and implementation can be observed.

Their weaknesses are equally instructive: attending a long meeting requires time, mobility, childcare, confidence speaking publicly, and sometimes tolerance for conflict. The formally open assembly may therefore be substantially less representative than the whole town.

Ireland: citizen formation followed by public authorization

Ireland’s recent constitutional processes are among the strongest modern examples of combining deliberation with referendums.

Randomly selected citizens heard evidence, questioned advocates, deliberated, and recommended constitutional changes. The abortion process subsequently led to the 2018 referendum in which 66.4% voted to remove the constitutional prohibition.

This arrangement divided political labor intelligently:

1. A representative mini-public studied a morally and technically complicated problem.
2. Parliament converted the recommendation into a legally operable proposal.
3. The full electorate made the constitutional decision.

However, Ireland’s rejected 2024 family and care referendums showed the fragile part of the chain: ambiguous legal wording and divergence between an assembly’s ideas and the eventual amendment can undo the legitimacy created by deliberation. The public must be able to trace the final ballot text back to the recommendations and see why changes were made.

Oregon’s Citizens’ Initiative Review

Oregon created a statutory process in which a demographically balanced panel of roughly 18–24 voters studies a ballot initiative for several days, questions advocates and experts, and prepares a citizen statement for the official voter pamphlet.

It does not let the panel write the initiative or make the final decision. Instead, it repairs the information stage by giving ordinary voters a concise assessment produced by other citizens rather than campaigns. The governing statute specifies balanced selection, compensation, minimum deliberation time, and presentation of arguments and findings. Oregon Revised Statutes §250

This is a modest but valuable model: attach a citizen jury to every major ballot proposition.

Ostbelgien: a permanent citizens’ council

The German-speaking Community of Belgium established a permanent system consisting of:

• A rotating 24-person Citizens’ Council selected by lot.
• Temporary citizens’ panels of approximately 25–50 people.
• Citizen-initiated topics.
• Parliamentary responses and public follow-up.

This goes beyond the one-off citizens’ assembly because citizens help set the continuing agenda and monitor what happens to recommendations. But Parliament retains final authority. The experiment is institutionally important, though its jurisdiction is small and long-term outcome evidence remains limited.

British Columbia: citizens develop an electoral system

A randomly selected citizens’ assembly studied electoral systems and proposed single transferable voting. In the 2005 referendum, its proposal received approximately 57.7% support and won in 77 of 79 districts—but failed because the government had imposed a 60% statewide threshold. It received much less support in a 2009 rerun. Official 2005 results

This demonstrates two things:

• Ordinary citizens can develop a sophisticated institutional proposal.
• A well-designed assembly does not ensure that millions of voters will understand or accept its work.

It also shows how a supermajority threshold can convert an actual majority into a formal defeat.

France’s Citizens’ Convention on Climate

France convened 150 randomly selected citizens for seven multi-day sessions, supported by experts and legal drafters. They produced 149 measures after receiving wider public contributions. Official convention record

The government had promised to transmit recommendations “without filter.” Many were subsequently altered, divided, or weakened. Different implementation trackers reached very different judgments about how many had been fulfilled.

The lesson is not that the assembly failed. It is that “government will consider the recommendations” is not an implementation rule. Before deliberation begins, participants need to know whether each recommendation will be:

• Enacted administratively.
• Introduced as legislation.
• Submitted to referendum.
• Returned with a public, reasoned rejection.
• Reconsidered by the assembly after modification.

Iceland’s crowdsourced constitution

Following the financial crisis, Iceland used a randomly selected national forum, an elected constitutional council, and online public input to draft a new constitution. In a 2012 advisory referendum, about two-thirds supported using the draft as the basis for a new constitution.

It was never enacted. The required parliamentary handoff was vulnerable to delay and political resistance. ConstitutionNet account

Its enduring lesson is stark:

Participation without an enforceable handoff can become elaborate consultation.

3. Do majorities inevitably become mobs?

No—but the danger is real under identifiable conditions.

A numerical loser is not automatically an oppressed minority. Every decision produces losers. “Tyranny of the majority” is more usefully reserved for decisions that deny a group equal legal status, fundamental rights, or meaningful future political participation.

Evidence that the risk is real

Barbara Gamble examined 74 US state and local initiatives concerning civil rights between 1959 and 1993. More than three-quarters of the measures restricting minority rights passed. “Putting Civil Rights to a Popular Vote”

Swiss naturalization provides even stronger quasi-experimental evidence. Some municipalities formerly decided individual naturalization applications by popular vote. After a court required elected councils to make these decisions with reasons and judicial review, naturalization rates rose by roughly 60%, especially for marginalized applicants and in more xenophobic municipalities. Hainmueller and Hangartner

This supports a strong design rule:

The public may legitimately vote on general immigration or citizenship rules, but it should not adjudicate the rights or status of named individuals.

Swiss research also finds that when visibly defined out-groups are at stake—especially religious minorities—direct votes and even the threat of referendums can push policy in a more restrictive direction. Christmann and Danaci

Evidence against the blanket “mob rule” claim

The broader proposition—that majorities invariably attack minorities—is not supported.

A study of 64 Swiss civil-rights-related votes classified 30% as producing anti-minority outcomes, with major variation by governmental level. That is troubling, but nowhere near inevitable, and it shows that institutional and political context matters. Frey and Goette

Moreover:

• Legislatures, courts, executives, and bureaucracies also have histories of discrimination.
• Electoral majorities are usually coalitions that change across issues.
• Referendums can overturn policies imposed by concentrated elites.
• The risk is different when voting on a reciprocal rule applying to everyone than when voting on the legal standing of an identifiable out-group.
• Deliberation, visibility, reasons, and judicial review change behavior.

The correct comparison is not “direct democracy versus perfect rights protection.” It is direct democracy versus the representative institutions that would otherwise decide.

The best protections

A durable system would combine popular sovereignty with:

• A universal constitutional rights floor.
• Judicial review governed by publicly stated principles.
• No popular adjudication of individual cases.
• General, reciprocal rules rather than group-specific punishments.
• A minority-impact statement before voting.
• A deliberative citizen review.
• Publicly disclosed campaign funding.
• A cooling-off period for constitutional changes.
• An opportunity to revise poorly drafted proposals.
• Equal access to a neutral voter guide.

Judicial review itself creates a democratic tradeoff: judges can protect rights but can also displace legitimate democratic decisions. The answer is a narrow, intelligible rights jurisdiction—not unlimited judicial supremacy.

4. What common claims survive the evidence?

5. When citizens help form ideas

The most interesting experiments do not begin with a finished ballot question.

Taiwan’s vTaiwan

vTaiwan uses a staged process involving proposals, public opinion mapping, reflection, stakeholder meetings, and governmental response.

Its best-known tool, Pol.is, allows participants to submit statements and agree, disagree, or pass on others’ statements. Instead of promoting the most inflammatory comment, it maps opinion clusters and highlights statements that receive support across otherwise opposed groups.

The platform reports more than two dozen cases and significant influence on digital-economy regulation. Those success figures are partly self-reported, so they should not be treated as independent causal evidence. vTaiwan process

Its main contribution is conceptual: look for cross-group agreement, not simply the loudest or largest faction.

Limitations include self-selection, facilitator power, a largely technology-oriented scope, and the fact that government usually retains final authority.

Estonia’s Rahvakogu, or People’s Assembly

Estonia’s 2013 process moved through several layers:

1. Open online submission of ideas and comments.
2. Expert consolidation and impact analysis.
3. Thematic workshops.
4. Deliberation by a recruited representative sample.
5. Transmission of proposals to Parliament.

The platform received thousands of ideas and comments. Fifteen proposals reached Parliament, and several produced legal changes, including reforms to the petition process. European Commission case study

This may be one of the closest prototypes for large-scale public idea formation. Its vulnerable point is the expert-consolidation stage: whoever groups, merges, and labels submissions exercises considerable agenda power.

A trustworthy system should therefore preserve:

• Every original submission.
• The reason proposals were merged.
• Traceable links from summaries to sources.
• Appeals against misclassification.
• Minority reports and alternative summaries.

Deliberative mini-publics generally

The OECD examined nearly 300 representative deliberative processes and identified 12 recurring models. It concludes that these processes are especially suitable for complex tradeoffs, long-term questions, and value conflicts. OECD report

Good processes normally include:

• Random or stratified selection.
• Payment, childcare, accessibility, and travel support.
• At least several full days of work.
• Balanced evidence and citizen ability to request additional experts.
• Skilled, independent facilitation.
• Public documentation of the evidence and method.
• A government obligation to answer every recommendation.
• Later reporting on implementation.

Random selection improves descriptive representation, but it is not magic. Nonresponse, expert framing, facilitator influence, group conformity, and small sample size remain concerns. Anonymous pre- and post-deliberation votes and explicit minority reports help.

6. Participatory budgeting

Participatory budgeting, pioneered in Porto Alegre in 1989, lets residents develop and select projects funded from a designated public budget. It has since appeared in thousands of cities. Participatory Budgeting Project overview

A typical cycle is:

1. Residents identify needs.
2. Participants develop project proposals.
3. Officials conduct legal and technical feasibility reviews.
4. Costs are published.
5. Residents vote.
6. Government funds winning projects.
7. A public tracker reports implementation.

Evidence of results

The strongest evidence comes from sustained Brazilian programs.

A study of Brazilian municipalities from 1990–2004 found that participatory budgeting shifted spending toward health and sanitation and was associated with a reduction in infant mortality of approximately 1–2 deaths per 1,000 births. Gonçalves

Research on major Brazilian cities likewise found associations with higher health spending, stronger civil society, and lower infant mortality, particularly where programs survived for multiple years and had support from the municipal government. Touchton and Wampler

But the evidence should not be oversold. Much of it is observational; municipalities that adopt and sustain participatory budgeting may already have stronger organizations or more capable governments. A public-health review found promising results but relatively few robust evaluations and only one randomized trial among the studies it examined. Scoping review

What makes it meaningful

Participatory budgeting becomes thin political theater when citizens control only a tiny collection of preapproved beautification projects.

Stronger versions provide:

• A consequential share of discretionary expenditure.
• Equity weighting based on poverty or unmet need.
• Open proposal workshops, not just online submission.
• Transparent feasibility criteria.
• Published capital and future operating costs.
• Assistance for communities with less grant-writing expertise.
• More than one voting method assessment.
• A binding implementation deadline.
• Public explanations for delay or cancellation.
• Repetition over many budget cycles.

The strongest distributive results appear where participatory budgeting is a durable institution connected to organized communities—not an annual app-based popularity contest.

7. The important technical choices

Binary voting

Simple majority is clean when exactly two clear alternatives exist. With three or more options, outcomes can depend on which pair is voted on first. Collective preferences can cycle:

• A majority prefers A to B.
• Another majority prefers B to C.
• Another majority prefers C to A.

That is why the agenda setter can become more powerful than the voters.

Possible responses include ranked voting, Condorcet methods, approval voting, or a two-stage process narrowing alternatives before final authorization. No method perfectly converts all individual rankings into one social ranking; the choice should be explicit and tested for sensitivity.

Quorums and thresholds

Turnout quorums encourage opponents to boycott rather than vote. Requirements that a proposal win support from a fixed percentage of the entire electorate effectively count every abstention as “no.”

Better approaches usually include:

• Ordinary majority for reversible ordinary policy.
• Higher thresholds for constitutional entrenchment.
• Two decisions separated by time for irreversible changes.
• A federal double majority only when territorial consent is genuinely relevant.
• Sunset and review clauses for experimental policy.

Thresholds should be symmetrical. It should not be easy to establish a restriction but nearly impossible to repeal it, or vice versa.

Ballot construction

Every proposition should have:

• A single-subject requirement.
• Independent legal review.
• A plain-language summary.
• A fiscal and administrative note.
• A rights and distributional-impact statement.
• The exact statutory text.
• A citizen-review statement.
• Visible identification of sponsors and major funders.

Participation versus representation

Three different signals should be published separately:

1. Open participation: what motivated people who chose to engage said.
2. Representative measurement: what a well-sampled population appears to prefer.
3. Deliberative judgment: what a representative group concluded after learning and discussion.

Comment counts are not opinion polls. Mini-publics are not elections. Elections do not show how preferences would change after study.

Digital security

Remote voting creates difficult conflicts among authentication, ballot secrecy, coercion resistance, device security, accessibility, and public auditing. A blockchain may preserve a recorded ballot but does not solve:

• Whether the voter’s device was compromised.
• Whether someone coerced the voter.
• Whether the voter sold credentials.
• Whether every eligible person received exactly one identity.
• Whether the software presented the same ballot it recorded.

High-stakes binding elections still benefit from a voter-verifiable physical record and risk-limiting audits. Digital platforms are currently more convincing for proposal development, discussion, delegation, and public tracking than as the sole election channel.

Artificial intelligence

AI can help cluster thousands of submissions, find duplicates, translate material, map arguments, and trace emerging areas of agreement.

It should not silently decide what “the public wants.” Required safeguards include:

• Source-to-summary provenance.
• Version histories.
• Multiple candidate summaries.
• Human appeal and correction.
• Publication of excluded material.
• Independent bias testing.
• Separation of moderation from political ranking.
• No conversion of engagement metrics into purported representative support.

8. Promising but under-tested models

• Liquid democracy: people vote directly or delegate their vote by topic, with revocable delegation. It could combine participation and expertise, but may create concentrated “super-delegates,” opaque influence networks, coercion problems, and unstable delegation chains. Large-scale binding evidence is sparse.

• A permanent citizens’ chamber: a randomly selected legislative chamber could review bills, launch inquiries, or delay legislation. Ostbelgien is a partial precedent, but a national chamber with binding veto or legislative power remains mostly theoretical.

• Quadratic voting: citizens receive equal voice credits but pay quadratically to cast multiple votes on one issue, revealing preference intensity. It may be useful for distributing a project budget, but complexity, collusion, identity verification, and strategic coordination remain concerns. Real money should not determine political voice.

• Deliberative polling: survey a representative group, expose it to balanced evidence and discussion, then survey it again. This is valuable for estimating informed opinion, although it generally has no formal authority.

• Citizen-triggered assemblies: a petition would trigger investigation by a random panel rather than immediately putting sponsor-written text on the ballot. This may offer a better balance between open agenda access and competent drafting.

• Adaptive legislation: citizens authorize a pilot with success metrics, followed by an audit and automatic reconsideration. This turns voting from a claim of certainty into a testable public decision.

• Futarchy: citizens or representatives specify social objectives while prediction markets estimate which policies would achieve them. The idea is intellectually interesting but faces serious problems in defining measurable values, preventing manipulation, and legitimizing decisions.

9. A high-performance democratic architecture

The most defensible synthesis of existing experiments would look like this:

1. Open agenda: anyone may submit or endorse a problem, not necessarily a finished law.
2. Transparent synthesis: proposals are clustered into a public issue tree, with full provenance and appeal rights.
3. Preliminary standing: a modest support threshold triggers study rather than immediate enactment.
4. Independent review: legal, fiscal, scientific, administrative, and rights implications are published; reviewers cannot silently kill a proposal.
5. Representative deliberation: a paid, stratified citizen panel hears competing evidence and develops alternatives.
6. Public revision: the wider population comments on the alternatives; the panel answers major objections and publishes minority reports.
7. Authorization: the electorate votes, preferably with more than a binary option when the issue genuinely has several alternatives.
8. Implementation: the relevant government must implement, seek a legally specified modification, or issue a reasoned refusal subject to review.
9. Audit: costs, outcomes, delays, and deviations appear on a public dashboard.
10. Reconsideration: experimental policies sunset or return for review after evidence accumulates.

That model combines different democratic strengths:

• Open participation finds neglected needs.
• Sortition supplies representativeness and concentrated learning.
• Experts contribute knowledge without owning the decision.
• Mass voting supplies authorization.
• Courts protect the constitutional floor.
• Administration supplies execution.
• Audits and revotes create learning.

The deepest lesson from Switzerland, Ireland, Brazil, Taiwan, Estonia, Iceland, Oregon, France, and Ostbelgien is that “the people” do not possess one preexisting will waiting to be counted. Public judgment is partly formed by institutions: what questions are visible, who participates, what evidence is available, which alternatives exist, and whether people believe their work will matter.

The real design question is therefore not simply “Should the majority rule?” It is:

How can a public discover problems, develop informed alternatives, authorize decisions, protect equal rights, and correct itself when it is wrong?

Yes—Me‑Us.

The architecture I described at the end is essentially the democratic operating system Me‑Us is building:

• Open comments surface what people care about.
• The issue tree organizes the problem without flattening it.
• Takeaways preserve distinct positions and their underlying reasons.
• Evidence attaches to claims.
• Sentence-level voting shows where agreement holds and where it breaks.
• Minority concerns become inputs for revising the proposal, rather than people to defeat.
• Distillation turns thousands of contributions into a coherent public judgment.
• Action tracking closes the gap between what the group decided and what institutions did.

Its most important difference from nearly every example above is continuity. Citizens’ assemblies, referendums, and participatory-budgeting rounds are temporary processes. Me‑Us could provide the persistent civic layer in which ideas are developed, tested, revised, authorized, and monitored.

The core claim becomes:

We have built systems for counting votes, but almost no infrastructure for forming an informed public will. Me‑Us is that missing infrastructure.

And the historical precedents divide neatly:

• Polis/vTaiwan: finding agreement across groups.
• Rahvakogu: turning mass input into organized proposals.
• Citizens’ assemblies: informed representative deliberation.
• Oregon CIR: citizen evaluation of finished proposals.
• Participatory budgeting: connecting public choice to material decisions.
• Switzerland: binding public authorization.
• Iceland and France: the warning—participation means little without an enforceable path to implementation.

The next useful pass would be to recast the whole deep dive as the evidence base for Me‑Us: what the precedents validate, what failures Me‑Us has to solve, and which product features follow from each finding.

I’m going to build this as the evidence case for Me‑Us and the Roundtable—not merely borrow democracy examples. I’ll first pull the relevant prior framing so the research tests your real claims: selective participation, modeled completion of partial views, facet-level affirmation, consensus formation, and dispersed control.

I found the current Roundtable specification and a Red‑Blue chapter among the relevant prior materials. I’m using those as the authoritative design baseline, because several of your strongest claims depend on mechanics already specified there—especially passive confirmation, human correction of inferred positions, provenance, and AI serving as facilitator rather than legitimacy source.

The underlying stored files are not directly openable in this session, but their current specifications and chapter findings were returned in enough detail to anchor the work. I’ll keep a hard line between what existing evidence supports, what it only suggests, and what remains a Me‑Us hypothesis that needs testing.

The evidence is splitting cleanly into three buckets: well-supported mechanisms (selective issue attention, structured deliberation, binding participatory budgets), conditional effects (legitimacy, depolarization, policy congruence), and claims that remain genuinely experimental (a three-answer civic model, reduced political burden, and displacement of the two-party system). I’m sharpening the answer around that distinction so the vision stays ambitious without turning inference into evidence.

Me‑Us as the missing civic layer

Yes—the pieces fit, and the distinctive idea is larger than “direct democracy.”

Me‑Us would treat democratic attention as scarce. It would not ask everyone to become a full-time legislator. A person could spend attention on the three or four things they genuinely care about, while an inspectable model develops a provisional picture of the rest. Roundtable then turns many such personal positions, arguments, questions, acceptable ranges, and lived experiences into coherent policy packages.

The essential sentence is:

The model does not replace my voice; it keeps my voice present when my attention is elsewhere.

A conventional election compresses a galaxy-shaped person into one red or blue bit. Me‑Us tries to reverse that compression without imposing hundreds of questions on every citizen.

The credible objective is therefore not permanent plebiscite. It is a continuous public-will infrastructure:

flowchart TD
A["Sparse personal input"] --> B["Editable personal policy vector"]
B --> C["Roundtable idea and concern map"]
C --> D["Evidence, questions, ranges, tradeoffs"]
D --> E["Facet-level consensus and packages"]
E --> F["Public authorization and implementation"]
F --> G["Outcome audit and revision"]
G --> C

This builds on the strongest existing Roundtable decisions: source provenance, direct correction by participants, signals counted by distinct people rather than comment volume, a living synthesis, and AI as facilitator—not as the source of legitimacy.

1. Selective interest is a feature

The premise that most people deeply care about only a few political questions is well supported.

Classic “issue public” research found that most people belong to only a few groups organized around highly salient concerns, shaped by their values, identities, and circumstances. On their important issues, their attitudes are more stable and politically consequential. Krosnick’s foundational study described a public composed of many small issue publics rather than uniformly attentive generalists. More recent work finds that accounting for differences in issue salience explains voting behavior better than simply sorting voters by political sophistication; conventional models understate issue voting when they assume every issue matters equally to everybody. Dufresne and Ouellet

That supports three Me‑Us ideas:

• Attention allocation is itself political information. Choosing housing, childcare, and privacy rather than foreign trade tells us something important about priority and intensity.

• Lived expertise can be more valuable than generic political knowledge. A disabled transit rider, renter facing eviction, farmer, teacher, or caregiver may understand particular policy failures better than a generalist representative.

• A citizen should not lose influence over everything else merely because they do not want to study everything else.

But missing input has two very different meanings:

1. “This is not important enough for my attention.”
2. “I lacked time, access, information, language support, or an invitation.”

Me‑Us must distinguish voluntary selectivity from exclusion.

It must also distinguish three outputs that ordinary participation platforms often blur:

• Participant view: What active contributors presently think.
• Population estimate: What a statistically representative public probably thinks.
• Authorized decision: What people have explicitly affirmed, voted for, or delegated.

A passionate self-selected group can develop the ideas. It cannot silently impersonate the population.

2. The three-answer personal model

Your strongest defensible claim is not “three answers tell us who someone is.”

It is:

Three answers may be enough to produce a useful draft whose errors are cheap to see and correct.

Sparse-preference recovery is technically plausible. Research using recommendation-system methods has shown that incomplete surveys can recover personal preferences, although the measurement format matters: continuous ratings produced stronger individual prediction at higher cognitive cost; binary choices were quick and useful for aggregate judgments but weaker for personal prediction; pairwise comparisons performed well for personal recovery but less well for aggregate judgments. Simple Surveys

But user-selected answers are not missing at random. People are likelier to rate things about which they have strong opinions. Collaborative-filtering research found that voluntarily supplied ratings differed substantially from randomly requested ratings, and prediction improved when the missingness mechanism itself was modeled. Marlin and colleagues

For Me‑Us, this means:

• Use chosen issues as evidence of salience.
• Do not assume unanswered issues resemble randomly omitted survey answers.
• Occasionally ask a small random “audit question” to measure inference error.
• Use representative samples—not merely open participation—to estimate the whole population.
• Never impute a nonparticipant’s binding vote.

A workable interaction

After three substantive inputs, Me‑Us might say:

“Here is your provisional policy profile. Twelve facets were inferred. Four are high-confidence, five moderate, and three uncertain. These two additional questions are most likely to change the resulting package.”

The next question should be chosen by something like:

That is better than marching everyone through the same 100-question questionnaire.

The model should be based primarily on issue relationships, prior answers, voluntarily disclosed circumstances, and lived constraints—not demographic stereotyping. “People like you” should mean people whose expressed tradeoffs resemble yours, not simply people of your race, ZIP code, age, or income.

And three models must remain separate:

• Preference model: What this person appears to value or accept.
• Outcome model: What various policies are predicted to cause.
• Consensus model: Which package can accommodate the group’s differing preferences.

If those are collapsed, the system can subtly convert an expert forecast into a supposed citizen preference.

3. Inference is not consent

Me‑Us needs an explicit authority ledger. Every stance should carry its provenance:

An overall affirmation can legitimately authorize the package without requiring 400 individual clicks. But the record must preserve that those facets were bundle-confirmed rather than individually considered.

I would draw a harder line around passive confirmation:

• In a discussion, failure to correct a visible synthesis can be weak evidence that it is serviceable.
• In a binding governmental decision, silence cannot become consent.

High-stakes, low-confidence, rights-affecting, or irreversible facets should require explicit attention. Material changes to a policy should expire authorization for the changed facets without forcing the person to reconfirm their entire political identity.

The model’s confidence must also be empirical—not an LLM saying it “feels 87% confident.” Calibration should be tested by hiding known answers, predicting them, and comparing predicted probabilities with actual responses across issues and population groups.

4. Roundtable is where public will is formed

Ordinary referenda ask citizens to accept or reject something already written by someone else. Roundtable’s deeper contribution is that people participate in forming the object being decided.

A Roundtable policy object should not be just a block of prose. It should be a versioned structure containing:

• problem definition;
• objectives and values;
• proposed policy levers;
• parameter values and acceptable ranges;
• affected groups and rights;
• costs and funding source;
• implementation responsibility;
• expected outcomes;
• uncertainties and model assumptions;
• conditions under which support changes;
• success measures, review date, and sunset provisions;
• source comments, evidence, objections, endorsements, and revision history.

The process then becomes:

1. People contribute proposals, experiences, concerns, or only endorsements and objections.
2. Roundtable extracts claims and places them in a nested idea taxonomy, always linked to the originals.
3. It distinguishes disagreement over values from disagreement over facts, forecasts, definitions, or implementation.
4. It identifies common unresolved questions.
5. Relevant experts answer those questions in a standardized, public format, including assumptions, uncertainty, conflicts of interest, and competing interpretations.
6. Participants react to those answers and may express conditional preferences: “I support this if cost remains below X” or “if the projected displacement rate stays below Y.”
7. Roundtable constructs several feasible packages rather than one artificial average.
8. People endorse or reject individual facets, acceptable ranges, and complete packages.
9. A rights review, legal review, budget check, and competing outcome models are applied.
10. The authorized package proceeds to implementation and public outcome auditing.

Consensus should be multidimensional

A bare “62% support” hides too much. Every significant facet should have a consensus passport showing:

• support;
• opposition;
• “can live with it” acceptability;
• salience and objection intensity;
• acceptable numerical range;
• number and representativeness of participants;
• direct versus bundle-confirmed authorization;
• uncertainty and unanswered questions;
• minority concerns and accommodations;
• stability after new evidence;
• rights and legal status;
• projected costs and outcomes.

The optimization goal is not unanimity. It is something closer to:

Find the feasible package with the broadest acceptability and lowest concentrated harm, subject to rights, legal, budgetary, and implementation constraints.

Consensus should shape the menu. Majority choice can settle what remains when a decision is unavoidable. Rights should constrain both.

5. What existing experiments establish

No existing case combines all the Me‑Us components. Each solves a different portion of the problem.

The comparative lesson is striking:

• Referenda provide power without enough public formation.
• Citizens’ assemblies provide formation without mass continuity.
• Digital consultations provide scale without secure representativeness or bindingness.
• Participatory budgeting provides bindingness but usually over a narrow slice of government.
• Me‑Us is attempting to join formation, scale, continuity, authorization, and implementation.

The OECD has catalogued nearly 300 representative deliberative processes and concludes that they are particularly suited to complex tradeoffs, long-term questions, and value conflicts. Its most important principle may be that the commissioning authority publicly commit in advance to responding, acting, and reporting on implementation. OECD

6. How probable are the larger outcomes?

7. What happens to Red–Blue politics?

Me‑Us can plausibly weaken the Red–Blue problem in three distinct ways.

First, it reveals that the coalitions are thin. A voter may vote Republican because of firearms and regulation while favoring Democratic positions on healthcare and labor. Another may vote Democratic because of abortion and climate while preferring conservative positions on zoning or education. Candidate voting hides the reasons, compromises, dislikes, and cross-party admiration.

Me‑Us routes the why.

Second, it creates different coalitions for different facets. The relevant grouping is no longer “Democrats versus Republicans,” but perhaps:

• renters and small developers versus incumbent homeowners on zoning;
• privacy advocates across parties on surveillance;
• rural patients and urban low-income patients on access;
• parents across ideologies on childcare affordability.

That can reduce identity sorting because the same two people may be allies on one issue and opponents on another.

Third, it can publish the true distribution before partisan interpretation takes over: “82% agree on the goal; disagreement concerns cost, timeline, and enforcement.” That is radically different from presenting every question as a red-versus-blue contest.

But the realistic endpoint is not a country without representatives. It is a country in which representatives no longer own the public’s voice.

Representatives, administrators, and experts would still be needed for:

• implementation;
• emergency action;
• diplomacy and negotiation;
• continuity and institutional memory;
• technical drafting;
• legal and constitutional review;
• accountability for operational failures.

Their role shifts from sovereign proxy to fiduciary implementer, facilitator, and publicly accountable steward.

8. Majority rule and “mob rule”

The fear of majority domination is real. Direct participation does not make 51% morally correct. Representative bodies are also capable of oppressing minorities, but moving the vote to the public does not remove the danger.

The Roundtable/Me‑Us response should be architectural:

• Constitutional floor: Equal standing, conscience, speech, religion, association, family, identity, due process, bodily integrity, and other basic rights are not ordinary preference variables.

• Different rules for different decisions: Routine reversible policies may use simple majority. Constitutional or irreversible decisions may require supermajorities, double majorities, multiple readings, cooling-off periods, or regional consent.

• Persistent-minority review: Track whether the same population repeatedly bears concentrated losses. Chronic losing should trigger accommodation, compensation, or rights review.

• Minority reports: Every synthesis must preserve the strongest unresolved objection and its evidentiary basis.

• No intensity multiplier on equal rights: Strength of feeling is useful for identifying where accommodation is needed; it should not automatically give some citizens more sovereign votes.

• Agenda transparency: Majority outcomes can be manipulated by the order and framing of alternatives. Who created, merged, excluded, or sequenced an option must be visible.

• Working consensus rather than unanimity: Requiring unanimity allows organized minorities to hold everything hostage. The objective is broad acceptability, preserved dissent, constitutional compliance, and a clear fallback decision rule.

A compact rule is:

Use deliberation to design; consensus to improve; majority rule to decide; constitutional rights to limit.

9. Corruption migrates unless every layer is dispersed

Dispersing the final vote makes one kind of capture harder, but it creates a larger attack surface. Influence can migrate to:

• which questions reach the agenda;
• how the taxonomy groups ideas;
• which assumptions appear as defaults;
• who is called an expert;
• which outcome model is displayed first;
• platform ownership and model training;
• paid signature or participation campaigns;
• targeted persuasion using private political profiles;
• implementation, procurement, and enforcement.

The political profile may become the most sensitive data a person owns. Nonnegotiable protections should include:

• separation of identity verification from political content;
• a secret-ballot authorization layer;
• user ownership, portability, deletion, and revocation;
• no political advertising or sale of preference data;
• public aggregation code and reproducible counts;
• multiple independently governed outcome models;
• signed expert evidence and conflict disclosures;
• independent security, bias, and implementation audits;
• offline, telephone, kiosk, disability, and language participation;
• strong one-person/one-credential protection for binding decisions;
• a public-interest or civic-trust governance structure for Me‑Us itself.

Otherwise Me‑Us risks becoming the new representative—an opaque platform standing between people and the state.

10. The experiment that would make this evidence rather than vision

A serious pilot should compare three conditions:

1. Conventional survey or comment forum.
2. Roundtable deliberation without a personal model.
3. Roundtable plus the sparse, editable Me‑Us policy model.

Give all three groups a real, binding budget allocation or institutional decision. Measure:

• personal-profile accuracy after 3, 5, and 10 answers;
• probability calibration on held-out questions;
• differences in error rates across groups;
• how often users correct the model;
• whether users understand which facets were inferred;
• time and cognitive burden;
• “this accurately represents me” ratings;
• perceived voice, fairness, and external efficacy;
• support, acceptability, and stability after evidence;
• quality of minority representation;
• partisan animosity and false beliefs about opponents;
• willingness to accept an outcome one opposed;
• congruence between public priorities and final expenditure;
• actual implementation, cost, delay, and outcome;
• attempts at organized capture or manipulation.

The decisive test is not whether the model guesses every answer. It is whether:

• it sharply lowers participation burden;
• users reliably detect consequential errors;
• inferred content never acquires unauthorized power;
• the resulting group policy is more faithful, coherent, legitimate, and implementable than the alternatives.

Bottom line

The evidence supports the possibility of Me‑Us more strongly than it supports the whole promised transformation.

The strongest supported chain is:

1. People really do concentrate concern on a few issues.
2. Sparse models can generate useful provisional preference drafts.
3. Structured deliberation can change views, surface common ground, and reduce some forms of polarization.
4. Participatory institutions can redirect real budgets and improve policy congruence.
5. Fair voice, correctability, and visible response can improve legitimacy.

The still-unproven chain is:

1. Three answers are sufficient across a full civic domain.
2. Model-assisted participation creates a durable feeling of agency while reducing political burden.
3. Facet-level common ground meaningfully displaces partisan identity at national scale.
4. Distributed participation substantially reduces corruption without equally strong audit institutions.
5. Me‑Us by itself retires the two-party system.

My own judgment is that the deepest opportunity is not replacing representatives with millions of legislators. It is separating functions that representative government has bundled together for centuries:

• citizens express ends, priorities, rights, and tolerable tradeoffs;
• Roundtable helps them form a coherent public will;
• experts model consequences and uncertainty;
• representatives and administrators execute;
• citizens retain authorization, correction, and audit power.

That is not government without representation. It is government in which representation ceases to mean surrender.