#1469 — Ranked Robin skips its own tiebreakers¶
Status: fix written, tested, committed locally. Not submitted — PR freeze.
Issue: #1469 (ours, filed 2026-08-01, two evidence comments)
Branch: fix/1469-ranked-robin-margins-tiebreaker @ 585b08f1, in the worktree /Volumes/T7/Voting/BetterVoting/bv-rr-degrees, off origin/main 454a38ae.
The defect¶
singleWinnerRankedRobin (packages/backend/src/Tabulators/RankedRobin.ts:32-72) resolves a tie for the most matchup wins with exactly two rungs: a head-to-head branch that fires only when exactly two candidates are tied, and then the random rung. Its own log line said so — "more robust tiebreaker not yet implemented".
Ranked Robin defines four degrees of tiebreakers (electowiki), and the first two are cheap:
- 1st Degree — elect the finalist with the greatest sum of win margins over the other finalists.
- 2nd Degree — still tied: greatest sum of win margins over all candidates.
The 3rd and 4th exist, and the spec explicitly does not recommend them for public elections — it prefers a lot or a re-run, which is what the random rung is. So the faithful ladder is 1st Degree → 2nd Degree → random, and the fix implements exactly that.
Two structural facts decide how much this matters:
- For exactly two finalists the 1st Degree IS the head-to-head, so the existing branch was already correct — it just had no name and no general case behind it.
- With three candidates and no drawn matchups, every Condorcet cycle is a three-way Copeland tie. The two-way branch can therefore never fire on a three-candidate cycle, and the random rung is the entire cycle path for the field size BetterVoting sees most. The issue's second comment carries the simulation.
Before → after¶
Both captured through packages/backend's own jest harness, not from a browser.
| Profile | On main |
With the fix |
|---|---|---|
| 11 ballots, 3-way cycle (the issue's repro: Dre 7–4 Edith, Edith 6–5 Frank, Frank 9–2 Dre) | Dre, tieBreakType: random, log "Dre picked in random tie-breaker, more robust tiebreaker not yet implemented." |
Frank, tieBreakType: none, log names the 1st Degree (+6 over the other finalists) |
| the same ballots, candidates listed Edith, Frank, Dre | Edith — the winner tracked the listing order | Frank |
| 26 ballots, 4 candidates, where the 1st and 2nd Degrees disagree | Ben (+24 over the field, almost all of it 26–0 against an also-ran) | Alma (+2 over the other finalists) |
| 4 ballots, two finalists drawn head-to-head | Alma by shuffle | Ben by the 2nd Degree (+4 vs +2 over the field) |
| 21 ballots, two finalists, decisive head-to-head | E | E — unchanged |
electowiki's "needs all four degrees" example (3r3yf7) |
random between Ava and Bianca | random between Ava and Bianca — unchanged, and tied now names just those two |
The last two rows are the guard: a fix that always decides would be the wrong fix.
Tests¶
packages/backend/src/Tabulators/RankedRobin.test.ts, six new cases, all expectations computed from the pairwise matrix by an independent stdlib script before the code was written.
- Baseline
npx jest src/Tabulators/: 9 suites, 52 tests green. - After: 9 suites, 58 tests green.
npx tsc --noEmit --project ./clean. - Reverting only
RankedRobin.tsand re-running: 4 of the 6 fail, the two regression cases pass. Sample failure —Expected: ["Frank"] / Received: ["Dre"].
Blast radius, stated honestly¶
tieBreakType stays 'none' on the deterministic rungs, so no results page starts announcing a tie where the method resolved one, and no election that already had a deterministic winner changes: the new rungs are reachable only from states that previously returned 'random'. The one other behaviour change is that roundResults.tied now lists the candidates the draw is actually between, rather than everyone who tied on matchup wins.
Not in the diff, deliberately:
- The log lines are plain English strings, matching the file's existing ones, not i18n keys. Worth raising with #1432 rather than solving here.
copelandScorecounts a pairwise draw as +0.5 while electowiki's primary rule says "beats the greatest number", which would count only outright wins. This can change who ties at the top. Pre-existing, out of scope, and not obviously a bug — the +½ convention is whatpref_votingand the LH engine also use, and wins+½·draws and wins−losses give identical rankings. It deserves its own issue with both readings quoted.- The sandbox assigns
tieBreakOrder = iwithout shuffling (sandboxController.ts), which is why the sandbox winner tracked the typed order. Unchanged here.
What this cost us, on our own side¶
Running the same ladder question against star-voting-library found the mirror-image bug in our engine: it had no 1st Degree rung at all and broke Copeland ties on total margin over the whole field, which is the 2nd Degree applied in place of the first. 11 of that repo's 100 Ranked Robin cases changed winner when it was corrected, every one a two-way tie whose head-to-head was decisive — i.e. the engine had been electing the loser of the finalists' own match.
Four of those cases are BV-backed, and in all four the corrected engine now agrees with the winner BetterVoting published. Two of them existed because of the disagreement: BV2138 and BV2176 were written up as an "LH vs BetterVoting tiebreak divergence". BetterVoting was right. Full write-up: degrees of ties.
That is worth a line in the PR body, and it is the argument for the PR: two independent implementations disagreeing is evidence that one is wrong, and here both were.
Credit¶
A concurrent session independently implemented the same ladder (branch fix/1469-ranked-robin-margins-tiebreakers @ 709d5c2e, kept in bv-copy-fix for salvage). Two things from it should be folded in before the PR opens: a 69-ballot regression test of the five-way cycle from the original Ranked Robin proposal, and printing the margin sums in the log lines so the export explains its own resolution.
Related¶
- #1063 — deterministic tie-breaking. This fix is most of it: after it, nearly every Ranked Robin tie resolves from the ballots.
- #1468 / #1480 — the results page stars a row by position, so it can disagree with the winner a tiebreak produced. Separate; PR #1479 is open on it.
- #1432 — surfacing tie-break explanations. The new log lines are raw material for it.