Methodology · Coordinationprovenverified

Voting-Based Group Decision

also known as social-choice voting methodology, preference-aggregation methodology

Applies to: multi-agent-system

Tags: social-choicevotingpreference-aggregationarrowgibbard-satterthwaite

Turn many agents' differing preferences into one group decision by choosing a voting rule that fits what the group needs. Different rules guarantee different things, such as picking a clear pairwise winner, rewarding honest votes, or dividing things fairly. The methodology is the procedure for choosing the rule, collecting votes in the right form, and admitting the limits no rule can escape (the Arrow and Gibbard-Satterthwaite results). It is more than just defaulting to majority vote. The rule choice is itself the design decision, and different rules can pick different winners from the very same votes.

Methodology process overview

flowchart TD prefs[Agent preferences + alternative set] --> rank[Rank desired axiomatic properties] rank --> elicit[Elicit in format the rule consumes] elicit --> pick{Pick voting rule} pick -- top-only --> plur[Plurality] pick -- full ranking --> borda[Borda count] pick -- pairwise winner --> cond[Condorcet method e.g. Copeland or Schulze] pick -- approval set --> appr[Approval voting] pick -- proportional --> stv[STV] plur --> tie[Specify tie-break] borda --> tie cond --> tie appr --> tie stv --> tie tie --> agg[Run aggregation on ballots] agg --> caveat[Document Gibbard-Satterthwaite manipulation surface] caveat --> audit[Audit ballots + outcome] audit --> out[Group decision + impossibility caveat]

Intent. Pick a voting rule whose guarantees fit the group decision, then combine the agents' votes under that rule while being clear about its limits.

When to apply. Use this when several agents hold different preferences over a fixed set of options and one group choice must be made. Examples are ranking tasks, choosing among options, voting on a config, or settling a multi-agent debate. Don't apply it when one agent's preference is authoritative; just defer to it. Skip it when preferences cannot be put into words; a numeric score may be more honest than a forced ranking. One exception: even with a single agent, a voting rule can combine several internal critics or model votes.

Example scenario

A platform group at a fintech runs a quarterly 'engineering priorities' vote across seven feature teams. The default had been plurality, where each team picks one favourite from the roadmap items. The winner was usually a splitting choice that nobody hated but few teams strongly wanted. After two quarters of dissatisfaction, the platform lead ran the voting-based group-decision methodology. They ranked the guarantees they cared about. They wanted to never pick an option when a unanimously better one exists. They wanted to pick the clear pairwise winner when one exists. And they wanted resistance to obvious manipulation. Plurality failed the pairwise-winner test badly on their past votes. Borda count failed it in one out of three replays, and the legal team flagged it as gameable by adding near-duplicate candidates. They picked the Schulze method, a rule that elects the pairwise winner and has a documented tie-break by path strength. For collection, each team submitted a full ranking of the 12 candidate items through a small internal app. The tie-break and manipulation caveats went into the announcement. What they learned: re-running past votes showed Schulze would have elected a different winner in two of the prior three quarters. After the fact, the teams preferred both of those winners. The first live Schulze election produced a clear pairwise winner. They documented the manipulation exposure honestly, since every non-dictatorial rule can be gamed somewhere (Gibbard-Satterthwaite). They named two specific manipulation patterns to watch for in audits. Different rules really do pick different winners from the same votes. Making the rule choice explicit moved the conversation from 'who won' to 'are we choosing well.'

Inputs

Agent preferences — Each agent's preference over the options. This may be a ranking, a score, an approval set, or pairwise comparisons, depending on how you collect it.
Alternative set — The fixed set of options being voted on, such as candidates, plans, tasks, or configurations.
Desired axiomatic properties — Which guarantees matter to you, ranked by what the deployment cannot live without. Examples include picking a clear pairwise winner, never punishing a candidate for gaining support, and resisting manipulation.

Outputs

Selected voting rule — The chosen rule, such as plurality, Borda count, approval voting, a Condorcet method, STV, or Copeland. The choice is justified by the guarantees you wanted.
Group decision — The single option, or the ranking, the rule produces from the agents' votes.
Impossibility caveat — A clear note of which guarantees the chosen rule gives up. Arrow's result proves you cannot have them all at once.

Steps (6)

Collect preferences in the right form
Rankings, scores, approval sets, and pairwise comparisons are not interchangeable. Collect votes in the form the chosen rule needs. If collection cost is the main concern, decide the form before you pick the rule.
Rank the guarantees you need
List the guarantees that matter, such as electing a clear pairwise winner, never punishing a candidate for gaining support, resisting manipulation, treating voters and candidates evenly, and respecting unanimity. Arrow's result says no rule can give all of them for three or more options. So rank yours and accept the losses you can live with.
Pick a rule whose guarantees match
Plurality is simple but ignores everything below each voter's top choice. Borda count uses full rankings but is easy to game. Condorcet methods, such as Copeland and Schulze, elect the pairwise winner when one exists. Approval voting reduces strategic guessing. STV gives proportional results.
Run the count
Apply the rule to the collected ballots. Some rules can hit a cycle where no clear winner exists (the Condorcet paradox). Specify the tie-break for that case and treat it as part of the rule, not an afterthought.
Document where the rule can be gamed
Every non-dictatorial rule over three or more options can be gamed by some voter in some case (Gibbard-Satterthwaite). Identify which of those tricks your deployment is exposed to. If the agents are strategic, this is a live risk.
Audit ballots and outcomes
Log every ballot and every count. Re-run the rule offline to confirm the result. Different rules can elect different winners on the same ballots, so an audit makes the rule choice visible after the fact.

Framework-specific instructions

Pick a framework and generate a framework-targeted rewrite of this methodology's steps.

Choose framework

AI-generated for Agent Development Kit (ADK) (Google) — verify against official docs.

Principles

The voting rule is the design decision. Defaulting to majority vote is choosing without choosing.
Arrow's result says you cannot have every good property at once. Rank them, accept the losses, and write them down.
The collection form and the rule must match. Use rankings for Borda and approvals for approval voting. Converting between forms loses information.
Every non-dictatorial rule can be gamed somewhere (Gibbard-Satterthwaite). Name where yours can.

Voting-Based Group Decision

Methodology process overview

Steps (6)

Collect preferences in the right form

Rank the guarantees you need

Pick a rule whose guarantees match

Run the count

Document where the rule can be gamed

Audit ballots and outcomes

Framework-specific instructions

Principles

Known failure modes (2)

Related patterns (3)

Sources (2)

Provenance