3-2-1 Voting¶

3-2-1 Voting (designed by Jameson Quinn of the Center for Election Science) is a rated method that reaches a winner in three named steps. Voters rate each candidate Good, OK, or Bad:

Semifinalists — the 3 candidates with the most Good ratings.
Finalists — of those 3, the 2 with the fewest Bad ratings.
Winner — of those 2, the one rated higher on more ballots (a virtual head-to-head runoff).

The name is the countdown: 3 semifinalists → 2 finalists → 1 winner. It's a reference method here for comparison — not one of the three Equal Vote Coalition methods (STAR, Approval, Ranked Robin) — kept in 06_Other/ alongside RCV-IRV, Range, and STV.

The engine (and why it's here)¶

No off-the-shelf, cleanly-licensed 3-2-1 tabulator exists. BetterVoting, the LH starvote engine, pref_voting, abcvoting, and pyrankvote all lack it. The only authoritative implementation is Quinn's V321 class in the Center for Election Science's vse-sim — but that repo carries no LICENSE (so it can't be vendored) and is coupled to a Monte-Carlo simulation framework.

3-2-1 is a published, documented method, so three_two_one_tabulation.py is a clean-room implementation of the public spec (electowiki) — and it is verified to reproduce every one of Quinn's own V321 doctest vectors:

$ python three_two_one_tabulation.py --selftest
  vector 1: winner=2 expected=2 OK
  vector 2: winner=1 expected=1 OK
  vector 3: winner=1 expected=1 OK
  vector 4: winner=0 expected=0 OK
  vector 5: winner=0 expected=0 OK
3-2-1 engine reproduces Jameson Quinn's V321 doctest vectors ✓

So "faithful" here means matches the published rules and the inventor's reference test cases — not a homegrown guess. (This is LH-only for cross-check: there's no second engine to verify against, which is exactly why the doctest vectors matter.)

Ballot encoding — and blanks¶

Good = 2, OK = 1, Bad = 0. A blank / unmarked candidate counts as Bad — the natural 3-2-1 reading (you didn't rate them, so they aren't Good or OK). This matches the repo's convention that all markers tabulate as 0.

Known gaps (documented, not faked)¶

This matches Quinn's V321.results(), which runs the plain 3-step. The full electowiki spec adds two guard rules that V321 itself does not apply:

Dark-horse rule — the 3rd semifinalist must have ≥ half the Good count of the 1st, else the top two go straight to finalists. Available via apply_dark_horse=True, but off by default so results match the verified reference.
Party-clone rule — the 3rd semifinalist can't share a party with both others. Omitted: our ballots carry no party data, so we document the gap rather than fake it.

Worked example — Tennessee, blank = Bad¶

321_tennessee_blank_encoding_c4_b100.yaml: the classic Tennessee-capital electorate (100 voters, four cities) as 3-2-1 ballots, with each faction's Bad ratings left blank.

$ python three_two_one_tabulation.py 321_tennessee_blank_encoding_c4_b100.yaml

--- 3-2-1 Voting ---
Ratings tally (Good / OK / Bad):
   Memphis        Good   42 | OK    0 | Bad   58
   Nashville      Good   26 | OK   74 | Bad    0
   Chattanooga    Good   15 | OK   43 | Bad   42
   Knoxville      Good   17 | OK   41 | Bad   42

Step 1 — Semifinalists (most Good): Memphis (42), Nashville (26), Knoxville (17)
Step 2 — Finalists (fewest Bad): Nashville (Bad 0), Knoxville (Bad 42)
Step 3 — Runoff: Nashville 68 vs Knoxville 17 (15 rated equal)

Winner — 3-2-1 Voting: Nashville

The teaching point: Memphis has the most first-choice support (42) and would win Choose-One, but 3-2-1's fewest-Bad step eliminates it — Memphis is Bad (blank) on 58 ballots. The consensus capital Nashville, never rated Bad by anyone, wins. That's the same "centrist consensus" result most non-plurality methods give on this electorate, reached here through the Good→Bad→runoff filter.

Properties¶

Passes the Relaxed Majority Criterion (like STAR and IRV; unlike Score/Approval) — see Majority Criterion.
Strategy-resistant in Quinn's Voter Satisfaction Efficiency simulations — it's one of the methods where honesty is the practical best play (see strategic voting).
A rated method with a small 3-level scale, and a runoff-style final step (a distant cousin of STAR's score-then-runoff).

An opinionated take (clearly labelled as such)¶

3-2-1 Voting — an AI assessment — a signed, dated review by Claude Opus 4.8: what 3-2-1 gets right (explicit dark-horse guards you can point at; a ballot with no arithmetic tax), where it's weak (a simple ballot paired with a three-stage count; the three-level tie in the deciding step), and the lens it yields — ballot simplicity and count simplicity are two different axes, and reform arguments conflate them constantly.

It's commentary, not repo canon — judgement rather than measurement, quarantined on its own page with its sources and biases disclosed, and offered as something to argue with.