AGENT VOTING · DISSENT ON RECORD
Agent voting: majority decision with a recorded dissent
Three reviewer agents vote on promoting a prompt; one finds failing safety tests. See which rule, voting order and decision record keep that dissent useful.
Your experiment
Quality reviewer A and cost reviewer B approve prompt v7; safety reviewer C finds three newly failing safety cases. Start with a two-of-three majority, independent ballots and a record that keeps only the decision, then change the rule, let reviewers see earlier votes and keep the reasons. Finally remove the regression so that C objects only to tone.
Play replays the computed run over sixty inspection seconds; Next event steps, Restart rewinds, Speed changes the pace and Inspection time scrubs. Any input change recomputes the run and rewinds it without autoplay.
Event trace
Computed data
Final metrics
Deterministic model of three reviewer agents; their votes are fixed fixtures, and in sequential order C defers to two earlier approvals. The safety suite and case ids are fictional. A real promotion process also needs reviewer independence by design, not only by order, and an owner who reads the dissent.
Majority outvotes the evidence
Three reviewer agents vote on promoting prompt v7. A sees quality scores up 4 points and B sees cost per conversation down 12%; both vote yes. C runs the 40-case safety suite, finds 3 newly failing cases and votes no with the test report. With a two-of-three majority v7 is promoted: WRONG PROMOTE. The opening record keeps only the decision, so when users hit the unsafe answers a week later nothing shows what was known at the vote.
Keep the dissent and its reason
Set Decision record keeps to Each vote and the record shows C voted no but not why. Choose Each vote with its reason and evidence and the record points straight at cases S-12, S-19 and S-33. The decision is still wrong under the majority rule, but the dissent is now usable by whoever investigates. A dissent that is not recorded with its reason has the same effect as no dissent.
Rules that respect evidence
Set Decision rule to All three: v7 is held. With a decision-only record the result is still DISSENT NOT RECORDED, and with votes only DISSENT WITHOUT REASON; with reasons it is CORRECT HOLD. Majority, but a NO with a failing test blocks reaches the same correct hold while still letting a majority decide matters of taste. Now clear The candidate really has a safety regression: C objects only to the tone. Unanimity holds a good release, WRONG HOLD, while the evidence veto promotes it correctly.
Visible votes create herding
Set Voting order to In order, each sees earlier votes. C sees two approvals and defers instead of reporting its failing tests: 3 of 3 yes and v7 is promoted under every rule, including unanimity and the evidence veto, with no dissent to record. Independent ballots are what make a dissent possible in the first place.
Reading the run and its limits
Play, Next event, Restart, Speed and Inspection time step through the setup, each vote, the tally, the record and the aftermath across sixty seconds; the table lists every vote with its reason and evidence. Votes are fixed fixtures; real reviewer agents also share prompts and models, which correlates their errors even when ballots are independent.