FairVote's single-winner comparison chart, claim-checked¶
FairVote's "Comparing single-winner voting methods" is the most-circulated method-comparison table in US reform — eight criteria, seven methods, a red/yellow/green grid. It is also the table a STAR advocate is most likely to be handed across a table. This page walks all eight criteria, quotes FairVote's own definition and justification for each, and marks it holds, overreaches, or false — with a runnable election wherever the claim is checkable.
Level: 301 · for debaters
The honest framing first. FairVote is the largest and oldest US reform organization and the chart is not junk: three of the eight rows hold as written, one of them grades FairVote's own method Low, and one leaves cells deliberately blank rather than guess. Where it fails it fails in a specific, diagnosable way — the rows are not graded by the same standard, and the standard slides in RCV's favor. That is the finding, and it is more useful than "FairVote is wrong."
Disclosure both ways. FairVote advocates RCV-IRV; this repo teaches STAR. Both leans are real. So every claim below is checked against a countable election or a published theorem, not against a preference — and three of the checks land against our own side and stay on the page.
→ The FairVote ledger this belongs to: FairVote's claims, checked · sibling claim-checks: the 2018 STAR white paper · the 2010 Condorcet article · our own grid, with the same warning label on it: Criteria at a glance
The verdicts, in one table¶
| # | FairVote's criterion | Verdict | The one-line reason |
|---|---|---|---|
| 1 | Well-tested in government elections | Holds | Accurate. But it measures adoption, not the method — the one row whose answer is guaranteed to favor whatever is already deployed. |
| 2 | Resistant to strategic voting | Overreaches | Later-no-harm and burial-immunity are real and conceded. "Strategic voting is not a concern" is not supportable — and the risk test applied to RCV's push-over is not applied to STAR's burial. |
| 3 | Resistance to "spoilers" | False | "RCV … satisfies … Independence of Irrelevant Alternatives" is false by Arrow's theorem and by two real elections — the same two elections FairVote names three rows later. |
| 4 | Majority cohesion | Holds, with a false consequence | STAR does fail mutual majority — conceded. But "vulnerable to the election of a candidate who lacks majority support" is wrong for STAR: its runoff is a head-to-head majority. |
| 5 | Condorcet efficiency | Mostly holds, one false sentence | The 2-in-~500 figure is real and we concede it. "STAR … fail[s] … the Condorcet Loser Criterion" is false — and FairVote's own white paper says so. |
| 6 | Simplicity of tabulation | Holds | Self-critical and correct. Missing the cost that actually bites: summability, where RCV and Condorcet methods are opposites, not equals. |
| 7 | Descriptive representation | Holds | Real research, correctly scoped, and untested methods get no rating rather than a red one. Good practice. |
| 8 | Compatibility with fair multi-winner | Holds, one false clause | STV's record is real and RCV's top mark is earned. "little study" of proportional approval/score is false — the literature starts in 1895. |
1. Well-tested in government elections — holds¶
"Has the method been tested in real, competitive elections for public office? … Condorcet methods, score, and STAR voting have never been used in a public election for government office, so any claims about their behavior in practice are unproven."
Conceded, without qualification. This is true, it matters, and STAR advocates should stop flinching at it. Jurisdictions genuinely do carry the risk of unintended consequences, and "we modelled it" is not the same evidence as "we ran it." FairVote's guinea-pig point is a fair description of a real political cost.
Two notes, neither a refutation:
- This row is a different kind of object from the other seven. Rows 3–5 are mathematical properties of a counting rule; this one is a fact about adoption history. In a chart whose purpose is to inform adoption decisions, a criterion that scores methods on having already been adopted is partly circular — every method now green was red on this row before its first election. That is not a reason to delete it. It is a reason not to average it with the others.
- One factual update, and it cuts FairVote's way. The text has approval voting "used in occasional municipal elections … with mixed success." That is now more pointed: Fargo adopted approval in 2018 and North Dakota banned it statewide in April 2025, ending it there; St. Louis's approval + top-two primary continues. A method's repeal list is half its record — this repo's approval page keeps both halves.
2. Resistant to strategic voting — overreaches¶
"RCV is most resistant to strategic manipulation and immune to the most common strategies: bullet-voting and burying. … As such, strategic voting is not a concern in jurisdictions and among voters that use RCV."
The mechanics are right. RCV-IRV satisfies later-no-harm, so adding a backup ranking cannot hurt your favorite and there is no incentive to truncate; and it is burial-immune, because your ballot only ever counts for your highest surviving choice, so demoting a rival buys nothing. Both are genuine strengths, and this repo says so on its own criteria grid.
FairVote's charge against STAR is also right, and we run it rather than argue it. STAR is vulnerable to burial, and the white paper's two worked examples — French 2017 and Washington 2010 — both succeed under STAR when we tabulate them: the burial cases, run.
The overreach is one sentence. "Strategic voting is not a concern" does not survive the paragraph it sits in, which already concedes that RCV-IRV is "vulnerable to compromising in rare circumstances." Compromising is favorite betrayal, and it is IRV's characteristic strategy precisely because IRV eliminates on first choices. It is not hypothetical: in Burlington 2009, Wright voters who ranked honestly got their last choice, and would have done better ranking Montroll first. A documented incentive in the most-studied RCV-IRV election in the country is a concern.
And the risk test is applied to one side only. FairVote dismisses the push-over strategy because it is "risky and difficult to pull off" and there is "no evidence of voters employing" it — a sound argument. The same argument applies to STAR's burial and is not made: our tabulation of FairVote's own French 2017 example shows the burial only works if every rival faction buries Macron and rates its ideological enemies a 4. Applied symmetrically, the test either excuses both or neither. (This repo's standing rule: apply the four-part test to both methods.)
Verdict: RCV's high mark on this row is defensible. The universal claim is not, and the asymmetric standard is the chart's first appearance of the pattern that breaks row 3.
3. Resistance to "spoilers" — false as stated¶
This is the row that fails, and it fails on its own terms.
"How well does the method prevent a minor candidate from causing a similar front-runner candidate to lose due to vote-splitting? Voting methods are resistant to 'spoilers' if adding or removing candidates … does not change the winner. … RCV is highly resistant to spoilers because it satisfies both the Independence of Irrelevant Alternatives and Independence of Clones criteria."
The theory says it cannot be true¶
Arrow's impossibility theorem — invoked by name in FairVote's own definition of this row — shows that no ranked rule over an unrestricted domain can combine IIA with Pareto and non-dictatorship. Arrow's own statement is about rules that output a full ranking; the single-winner analogue is Muller–Satterthwaite, and it lands in the same place. Either way RCV-IRV, a ranked rule and not a dictatorship, is squarely inside the scope. So "RCV satisfies IIA" is not a claim that happens to be false — it is a claim that cannot be true of any ranked method, asserted two sentences after citing the theorem that forbids it.
(This repo is picky about Arrow's scope precisely because comparison charts blur it — which criteria are actually Arrow's, and which are not. And the reason it matters here: rated methods are outside the theorem entirely, which is where the approval/score half of this row goes wrong. Arrow and STAR.)
And it fails on real ballots — twice¶
FairVote's own definition is a test you can run: delete a candidate who did not win, change nothing else, see if the winner changes. We ran it on the two most-studied RCV elections in the United States.
Alaska 2022 US House special — the 200-voter reduced model of the certified profile (Graham-Squire & McCune, Table 1). Delete Palin, who lost:
| First choices | RCV-IRV winner | |
|---|---|---|
| All three candidates | Peltola 80 · Palin 63 · Begich 57 | Peltola (96–92 over Palin) |
| Palin deleted from every ballot | Begich 93 · Peltola 84 (after Palin's removal) | Begich (93–84) |
--- RCV / Instant-Runoff Voting (single winner) ---
Alaska 2022 special (reduced model), Palin deleted from every ballot — RCV-IRV now elects Begich
Tabulating 177 ballots (ranked ballots).
FINAL RESULT
Candidate Votes Status
----------- ------- --------
Begich 93 Elected
Peltola 84 Rejected
Winner(s) — RCV / Instant-Runoff Voting (single winner)
Begich
Burlington 2009 mayor — the real 8,980-ballot record (PrefLib 00005). Delete Wright, who lost:
| RCV-IRV winner | |
|---|---|
| The certified election | Kiss (4,313–4,059 over Wright) |
| Wright deleted from every ballot | Montroll (4,063–3,476 over Kiss) |
--- RCV / Instant-Runoff Voting (single winner) ---
Burlington 2009 mayor, Wright deleted from every ballot — RCV-IRV now elects Montroll
Tabulating 8134 ballots (ranked ballots).
ROUND 1
Candidate Votes Status
----------- ------- --------
Montroll 2911 Hopeful
Kiss 2866 Hopeful
Smith 2188 Rejected
Simpson 113 Rejected
WriteIn 56 Rejected
FINAL RESULT
Candidate Votes Status
----------- ------- --------
Montroll 4063 Elected
Kiss 3476 Rejected
Smith 0 Rejected
Simpson 0 Rejected
WriteIn 0 Rejected
Blank Votes 595 Rejected
Winner(s) — RCV / Instant-Runoff Voting (single winner)
Montroll
--- Transfers and inactive ballots (what the round tables leave out) ---
The tables above give each candidate's round total but not where a
transferred vote came FROM, nor how many ballots stopped counting.
Both are recomputed from the ballots, using the eliminations the
count above actually made.
ROUND 1 — 8134 of 8134 ballots still active; majority = 4068
WriteIn eliminated with 56:
→ (no continuing ranking) 26 ← these ballots go inactive
→ Montroll 22
→ Kiss 8
Simpson eliminated with 113:
→ (no continuing ranking) 44 ← these ballots go inactive
→ Montroll 35
→ Kiss 34
Smith eliminated with 2188:
→ Montroll 1095
→ Kiss 568
→ (no continuing ranking) 525 ← these ballots go inactive
FINAL ROUND — 7539 of 8134 ballots still active (595 inactive); majority = 3770
Montroll 4063 (53.9% of the still-active) ← elected
Kiss 3476 (46.1% of the still-active)
Never exhausted, never transferred:
2884 ballots held by Kiss carried a lower ranking that was never read
(the count stopped here, so those preferences did nothing).
Inactive ballots at the final round: 595 of 8134 (7.3%).
Montroll's 4063 is a majority of the 7539 still active but only 50.0% of all 8134 cast —
the 'majority' here is of a shrunken denominator. See
06_Other/RCV_IRV/concepts/RCV_IRV_exhausted_ballots.md
A built-in check that these are not artifacts. In both counterfactuals the final round reproduces, digit for digit, a head-to-head the full election's own pairwise matrix already reported — Begich over Peltola 93–84, Montroll over Kiss 4,063–3,476. Removing the extra candidate does not create a new election; it reduces the instant runoff to a matchup the ballots always contained and the count in force never asked about.
What the counterfactual does and does not assume. Ballots are held fixed and one candidate is struck out; no voter is handed a preference they did not express, and ballots left empty (840 in Burlington, 23 in Alaska) simply exhaust. A real withdrawal would also have moved campaigns and turnout — but IIA is defined on fixed ballots, so this is the criterion's own test, and it is the only version a certified ballot record can honestly support.
Which criterion is actually satisfied — and it is not the one being described¶
FairVote names two criteria. They do not have the same status here:
- Independence of clones — genuinely satisfied. RCV-IRV passes it, and Alaska does not violate it: 4 ballots rank Peltola between the two Republicans, so {Palin, Begich} is not a clone set in the technical sense. Credit where due.
- Independence of Irrelevant Alternatives — failed, on both elections above.
The gap matters because IIA is the criterion FairVote's own prose defines ("adding or removing candidates … does not change the winner"). The chart's plain-English test is IIA; the criterion RCV actually passes is the narrower one; and the two are presented as if RCV had both.
The internal contradiction¶
Three rows later, on Condorcet efficiency, the same page states that of "nearly 500" US RCV elections with full ballot data, the beats-all winner "only lost twice" — and FairVote's own data page names those two: Burlington 2009 and the Alaska 2022 special. They are the two elections above. A method that satisfied IIA could not have produced either. The chart grades RCV High on spoiler resistance and Medium on Condorcet efficiency using the same two elections as evidence in opposite directions.
The double standard on approval and score¶
"Both approval voting and score voting are more resistant to spoilers than plurality voting … However, the expectation that voters will behave in this fashion depends on three assumptions … voters need to know who the front-runners are … there must only be two clear frontrunners … voters must be comfortable insincerely giving a front-runner the same score as their actual favorite."
Every one of the three assumptions is about voters changing their ballots. But the row's own definition — and the standard applied to RCV one paragraph earlier — holds ballots fixed. Under that standard, score voting passes IIA outright: deleting a losing candidate cannot change any other candidate's total, because each candidate's score is independent. Approval likewise. Cardinal methods are not subject to Arrow at all, which is the whole reason the theorem's scope is worth knowing.
The genuine kernel, conceded. If voters normalize — rescaling 0–5 around whoever is running rather than on an absolute scale — then the ballots do change when a candidate enters, and a residual spoiler returns. That is real, and this repo documents it under its own name: STAR's residual vote-splitting. FairVote's Medium for approval/score is therefore a defensible behavioral prediction. It is not defensible as an application of the row's stated test, which grades RCV by a rule it then abandons.
STAR: right rating, wrong axis¶
"STAR voting is more resistant to spoilers than plurality voting, approval, or score voting but is still vulnerable to spoilers due to its susceptibility to strategic voting in the form of 'burying'."
Burying is a strategy, and strategy has its own row two criteria earlier. Grading STAR down here for it double-counts, and STAR is the only method in the row graded on strategy rather than on candidate entry and exit.
STAR does fail this row — but for the reason FairVote does not give. Under a genuine Condorcet cycle, a candidate who cannot win can still change which two candidates reach the automatic runoff, on perfectly sincere ballots and with no strategy at all: STAR's IIA limit, run (BV2212). So STAR earns its Medium honestly, by a mechanism the chart never mentions.
The row, graded by its own definition¶
Applying FairVote's stated test — fixed ballots, add or remove a candidate — consistently across all seven columns:
| Method | FairVote | By FairVote's own test | Why |
|---|---|---|---|
| Plurality | Low | Low | Agreed — the textbook case. |
| Two-round runoff | Medium | Medium | Agreed, and FairVote's third-place example is a good one. |
| RCV-IRV | High | Medium | Passes clones, fails IIA — Burlington 2009, Alaska 2022. |
| Approval | Medium | High* | Passes IIA on fixed ballots. *Medium if voters normalize to the field. |
| Score / Range | Medium | High* | Same — and cardinal methods are outside Arrow's scope. |
| STAR | Medium | Medium | Correct rating; the reason is the cycle, not burial. |
| Condorcet methods (Ranked Robin) | High | High | Agreed — fails only inside a cycle. |
The ordering the row's own definition produces is close to the reverse of the one printed for the top three.
Where the wider treatment lives: the spoiler effect · how often vote splitting actually happens · the runnable split-voting progression.
4. Majority cohesion — holds, with a false consequence¶
"RCV is perfect in this regard: It satisfies the Mutual Majority Criterion … Approval, score, and STAR voting do not satisfy either criteria related to majority cohesion. These methods are vulnerable to the election of a candidate who lacks majority support."
Conceded on the criteria. STAR fails majority favorite and mutual majority in constructed cases; RCV-IRV passes both, and so does Ranked Robin. This repo's own grid says the same, and its honest-limits page does not hedge it.
The consequence sentence is wrong for STAR. STAR's automatic runoff is a head-to-head between the top two scorers, so the winner always beats the runner-up among voters expressing a preference (bar an exact runoff tie, which the score round breaks — the one seam, and it is why STAR fails the weak Condorcet-loser criterion: a tie is not a loss). STAR can elect someone who was not a majority's first choice — that is the majority-favorite failure, and it is real — but "lacks majority support" describes plurality, which is graded Low on this row and correctly so. The two failures are not the same thing, and collapsing them is what makes the sentence false rather than merely unflattering.
The row that is missing. A chart with a majority-cohesion criterion and no exhausted-ballot row leaves out the majority complaint voters actually make. RCV-IRV's reported majority is a majority of continuing ballots, not of ballots cast — in the Alaska 2022 reduced model, Peltola's winning 96 is a majority of the 188 still-active ballots and only 48% of the 200 cast (the round-by-round math). That is why this repo's engine prints a transfer block reconciling the final round against all ballots cast, and why false majorities is its own page.
5. Condorcet efficiency — mostly holds; one sentence is false¶
"Of the nearly 500 RCV elections in the United States since 2004 for which full ranked-ballot data are available, the 'beats-all' winner only lost twice — a Condorcet efficiency rate of 99.6% in practice."
Conceded, and loudly. This is a real, checkable, important number, and it is the single best argument on the page. Most RCV elections work fine. A STAR advocate who implies otherwise is misleading people, and this repo's rule is to state the rarity every time it criticizes IRV.
The false sentence:
"Plurality, approval, score, and STAR voting fail the Condorcet Criterion, but they also fail a far weaker property known as the Condorcet Loser Criterion."
STAR passes the Condorcet loser criterion. A candidate who loses every head-to-head can top the scoring round, but they lose the automatic runoff by construction — that is what the runoff is for. And FairVote knows this: its own 2018 white paper concedes that STAR cannot elect the Condorcet loser, a point we quote and credit. Two FairVote documents contradict each other, and the chart is the one that is wrong. (Two-round runoff passes it too, and is not in the list.)
The framing hides the field-size effect. The 99.6% is an average over a corpus dominated by small, low-competition fields; it is a statement about the elections RCV has been used in, not about the rule. Measured against field size — our simulation, 4,000 elections per cell, model and seed published — RCV-IRV's Condorcet efficiency falls from ~96% at three candidates to ~47–52% at seven in a one-dimensional spatial electorate, while STAR holds 74–92% and Ranked Robin is 100% by construction. The failures concentrate in exactly the crowded, competitive races reform is for: why more candidates miss.
Fairness cut, stated plainly: ours is a model and FairVote's is real election data. Real data wins on what has happened; the model speaks to what happens as fields grow. The honest synthesis is that RCV's practical Condorcet record is good, and it degrades where the stakes are highest — which is also why the two failures are Burlington and Alaska rather than two sleepy school-board races.
6. Simplicity of tabulation — holds¶
"RCV and Condorcet methods are more complex than a simple arithmetic sum, and are therefore harder to explain and implement."
Correct, and worth pausing on: FairVote grades its own method Low here. That is what good-faith comparison looks like, and it deserves to be said as loudly as the criticisms above.
One thing the row omits, and it is the cost that actually bites. The administrative problem with IRV is not arithmetic difficulty — it is summability. An IRV count cannot be assembled from precinct subtotals; every ballot has to reach one place before any round can be run, which is what drives central tabulation, delayed results, and harder recounts. Condorcet methods are summable — each precinct reports an N×N pairwise matrix and the matrices add. Score and STAR are summable too (one running total per candidate). Grading RCV and Condorcet methods the same red treats opposites as equals on the property election administrators care most about.
7. Descriptive representation — holds¶
"RCV has demonstrably improved representation for women and people of color. … Approval, score, STAR, and Condorcet methods are untested in practice. No evidence shows these methods would improve the diversity of our elected representatives."
Holds. The research is real, the claim is scoped to it, and — notably — the untested methods are left unrated rather than marked red. Leaving a cell blank when the answer is unknown rather than bad is better practice than most comparison charts manage, including some pro-STAR ones.
Two caveats, neither fatal. Much of this literature is produced by FairVote and allied researchers, which is a lean worth disclosing the same way this repo discloses its own; and a meaningful share of the effect is attributed to multi-winner RCV and to candidate emergence rather than to the single-winner tabulation this chart is about. Both are reasons to read the row carefully, not to discount it.
8. Compatibility with fair multi-winner elections — holds, one false clause¶
"RCV earns a top score in this area because its multiwinner form, proportional RCV (aka the Single Transferable Vote), is an accepted and well-tested method … While some advocates have proposed proportional analogs to Condorcet, approval, score, and STAR voting, they have seen scant or non-existent use and little study or advocacy."
The core point holds and the top mark is earned. STV has more than a century of government use, and running single- and multi-winner offices off one ballot type is a real practical advantage. Conceded.
"Little study" is false. Proportional approval voting is Thiele, 1895; Phragmén's methods are the same decade; the modern literature has its own textbook — Lackner & Skowron, Multi-Winner Voting with Approval Preferences (Springer, 2023, open access) — and a maintained reference implementation, abcvoting, which this repo runs as a cross-check engine on its own approval cases. Whatever else is true of these rules, "little study" is not.
"Scant or non-existent use" is closer to right, and that is the honest core — for government elections it is essentially correct, and FairVote is entitled to the point. Outside government it is not: the Academy of Motion Picture Arts and Sciences uses reweighted range voting every year to cut ten shortlisted films to five Best Visual Effects nominees (Academy press office, Variety) — a proportional score method running a high-stakes annual election. The repo's runnable side: STAR-PR · STV vs STAR-PR on one electorate · comparing multi-winner methods.
The structural finding¶
The rows are three different kinds of object, and the grid renders them identically. Row 1 is a deployment fact; rows 3–5 and 8 are mathematical properties; rows 2 and 7 are empirical predictions about behavior. A red cell in row 1 means "nobody has tried it"; a red cell in row 3 means "provably fails." Same color, incomparable content — and a reader scanning columns for green will add them up.
The standard slides. Row 3 grades RCV on fixed ballots and cardinal methods on hypothetical voter behavior. Row 2 excuses RCV's push-over strategy as too risky and coordination-heavy to matter, and does not extend that reasoning to STAR's burial. Row 5 states a criterion failure for STAR that FairVote's own white paper elsewhere concedes STAR passes. Any one of these is an ordinary advocacy slip; together they are a direction.
And the same warning applies to our grid. This repo publishes its own criteria table with the caveat printed above it: a pass/fail chart implies every criterion matters equally and that "fails" is binary, when the real questions are how often and how badly. Even STAR's own advocates argue against the format ("Farewell to Pass/Fail"). Use FairVote's chart — and ours — to navigate to the worked elections, not to crown a winner.
Run it yourself¶
.venv/bin/python STARVote_LH_tabulation_engine/starvote_larry_hastings.py method_comparisons/alaska_2022/cases/alaska_2022_irv_with_palin.yaml
.venv/bin/python STARVote_LH_tabulation_engine/starvote_larry_hastings.py method_comparisons/alaska_2022/cases/alaska_2022_irv_without_palin.yaml
.venv/bin/python STARVote_LH_tabulation_engine/starvote_larry_hastings.py method_comparisons/burlington_2009/cases/burlington_2009_irv_without_wright.yaml
Sources. FairVote, "Comparing single-winner voting methods" (quotations retrieved 2026-08-22) and "Research and data on RCV in practice" (the two Condorcet failures, named). Ballot data: PrefLib 00005 (Burlington 2009); Graham-Squire & McCune, An Examination of Ranked Choice Voting in the United States, 2004–2022 (arXiv:2301.12075, Table 1 — the Alaska profile). Neutral references: Independence of irrelevant alternatives · Arrow's impossibility theorem · 2009 Burlington mayoral election.