Skip to content

The Ranked Robin Whoops Files 🍿 — mud for the consensus darling too

⚠️ Deliberately unfair, and it's about a method this repo likes. Ranked Robin (Condorcet/Copeland) is the library's olive branch to ranked-choice voters — same ballot, a monotonic, summable, center-squeeze-free count. So naturally it gets the same mud-bucket as everyone else in the Mudroom. Every "whoops" is real and conceded in RR's honest limits; the curation is the unfair part.

Condorcet methods look unbeatable — until the electorate refuses to have a "beats-everyone" candidate at all.

🥇 The signature flaw: cycles

  • No Condorcet winner exists. A majority prefers A>B, B>C, and C>A — rock-paper-scissors, so "elect the head-to-head winner" has nothing to elect. Ranked Robin falls back on a tiebreak (most wins → 1st Degree → 2nd Degree → lot), which is a choice, not a discovery. → cycle resolution; a live one in the reversal-symmetry case (that 24-voter electorate is a cycle). This is Gibbard collecting his due — no method escapes it, RR included.

🥈 Order without strength

  • The sincere dark horse. Because a ranked ballot sees order, not intensity, RR can crown a thinly-supported unknown everyone merely tolerates as a second choice over a candidate a majority passionately prefers — "somebody 90% of voters never heard of," each side ranking them above the other side. → RR's honest limits and the sincere dark horse. (This is exactly the preference-vs-support blind spot — the one thing STAR reads and RR can't.)

🥉 The bureaucratic whoops

  • Tie-broken by lot — and nobody implemented the ladder. Ranked Robin ships with a four-rung tie-break protocol, and for two years neither engine in this repo ran it: BetterVoting handles only a two-way tie and shuffles anything bigger, while this library's own engine skipped straight to the whole-field margin rung and could hand the race to a candidate who lost the finalists' own match. Both are now known bugs rather than rival conventions — ours fixed on 2026-08-19, BV's filed as #1469 — but the mud stands: a method whose tie-break is subtle enough that two independent implementations both got it wrong is a method whose knife-edge races were decided by which counter you used. → degrees of ties.
  • Copeland's clone wrinkles. Simple win-counting can be nudged by adding clone candidates in some constructions. → RR clone independence.

The conscience clause

The fair version, because the Mudroom insists:

  • Cycles are genuinely rare (~1–5% in realistic models and empirical data), and every method must resolve them somehow — RR is not uniquely cursed here.
  • RR avoids the failures that actually bite in real elections — it has no center squeeze and no non-monotonicity, which is more than IRV can say.
  • The tie-break mud lands on the counters, not on the method. Ranked Robin's degrees of ties were published, worked and four rungs deep the whole time; two implementations failed to follow them, in opposite directions. That is an argument for reading the spec, not against the method — and this repo's engine was one of the two, which is why the story is told here rather than filed under someone else's failures.
  • The intensity-blindness is the honest cost of a ranked ballot; it's why STAR pairs scores with a runoff — a real distinction, not a knock that makes RR "bad."
  • RR remains a strong, summable, honest method and the repo's genuine bridge to RCV voters. Mud thrown; respect intact.

For the version you'd actually cite, use RR's honest limits.