3-2-1 Voting — an AI assessment¶
An opinionated review of 3-2-1 voting written by Claude Opus 4.8 on 2026-07-23, at Adam's request. This page is commentary, not repo canon — it contains judgements, not measurements. Everything factual in it is sourced below; everything evaluative is one model's read and should be argued with.
How to read this page. The rest of the library states things it can verify — a tabulation, a criterion, a simulated rate with its model attached. This page does something different: it judges. Judgements can't be regression-tested, so they're quarantined here, signed and dated, rather than mixed into the concept pages. If you want the neutral treatment, read 3-2-1 voting and the method folder instead.
Provenance and lean (read first)¶
- What I read: this repo's 3-2-1 concept page, the folder README with Quinn's spec, and the clean-room engine verified against Quinn's own reference vectors.
- What I did not read: the electowiki 3-2-1 article — it returned HTTP 403 when fetched. That is the method's primary public definition and it may raise criticisms I haven't weighed. This is a real gap in the review.
- My lean: I formed this view largely from pages written in a STAR-advocating house style, so it likely under-sells 3-2-1. I've tried to correct for that and probably haven't fully.
- Bottom line I'd defend: 3-2-1 is a good method; I'd take it over RCV-IRV without hesitation; I'd take STAR over it, but I hold that last preference loosely.
What's genuinely good¶
The explicit guards are underrated. The ½-Good dark-horse rule and the fewest-Bad finalist step mean that when someone asks "what stops a nobody from winning?", you point at a rule. STAR answers the same question correctly but emergently — the scores simply happen to prevent it. In a legislative hearing, a rule you can point at beats a property you have to derive. That is a real political asset, and STAR advocates tend to treat it as a wart.
Good / OK / Bad is more legible than 0–5 to a genuinely skeptical audience. Three words, no arithmetic implied. I think the STAR world underestimates how much "score each candidate 0 to 5" reads as homework to people who aren't already interested in voting theory. 3-2-1's ballot has no such tax.
The outcomes are fine. No center squeeze, a non-slippery chicken dilemma, summable, high VSE. It's in STAR's league, and both are in a different league from IRV and plurality.
Where I think it's wrong¶
1. It simplifies the wrong end¶
3-2-1 pairs a simple ballot with a three-stage count containing special cases — the ½-Good threshold, the skip-step-2 conditional. STAR trades the other way: a slightly richer ballot, then add the scores and hold one runoff.
Voters touch the ballot once, in private, with instructions in front of them. Officials, auditors, journalists, judges and opponents touch the count forever. And the count is precisely where RCV-IRV bleeds — summability, central tabulation, "explain to me why my candidate was eliminated."
Optimizing ballot simplicity while adding count complexity is spending your simplicity budget on the easy problem.
2. The granularity tie is structural, not an edge case¶
Three levels means the deciding step often has nothing to read. If two allies are both honestly Good, no voter distinguishes them, and step 3 is a coin flip — the repo's own worked example shows exactly this.
STAR's fix is free: scoring your favorite 5 and your ally 4 is natural and costs nothing emotionally. 3-2-1 needs voters to spend the OK level strategically — to demote someone they actually like — for the final step to function. That's a small honesty tax baked into the design, and it lands on exactly the cooperative voters a good method should protect.
3. Bolt-on constants age badly¶
"At least ½ as many Good ratings" invites the obvious question: why one half? There is no principled answer, only a tuned one. Every such constant is a surface for an opponent to attack, and a thing that can be wrong for a particular electorate. I'd rather have properties that fall out of the mechanism than parameters someone chose.
(To be fair to Quinn: he was a statistician who tuned constants against simulation, which is a better reason than intuition. The objection is about political durability, not rigor.)
The strategic reality¶
3-2-1 is an orphan. Quinn designed it and died in March 2025; the Center for Election Science campaigns for Approval, the Equal Vote Coalition for STAR and Ranked Robin. Nobody is carrying it.
That is not a criticism of the method — plenty of good ideas lack a campaign — but it means the practical choice essentially never comes down to "STAR or 3-2-1." Its live role is as a reference point that sharpens how you argue for the others, which is exactly why this repo keeps a working engine for it.
What I actually take from it¶
The most valuable thing about 3-2-1 isn't the method. It's that 3-2-1 is the cleanest illustration that "simplicity" is two axes, not one — a simple ballot bolted to a comparatively involved count. The neutral treatment, with every method placed on both axes, is What makes a voting method good? §2 (see also Is RCV "simple"? — which half?).
Once you hold the axes apart, the STAR-vs-3-2-1 disagreement stops being "which is simpler?" — unanswerable, since each side points at a different axis — and becomes "which kind of simplicity is worth more?", which you can actually argue.
My answer: count simplicity is worth more, because it's the axis that determines whether officials can administer the method, auditors can verify it, and opponents can misrepresent it. Ballot complexity is a one-time explanation cost. Count complexity is a permanent attack surface.
Reasonable people land differently on that, and if you think voter comprehension is the binding constraint on adoption, 3-2-1 beats STAR on your own terms.
Written by Claude Opus 4.8, 2026-07-23. Opinion, openly biased toward the repo's house methods, and offered as something to disagree with. Corrections and rebuttals are more useful to the library than agreement — if a claim here is wrong, fix the page rather than preserving it.