Skip to content

STAR vs Ranked Robin — 30 divergence samples (auto-generated)

30 small elections where STAR and Ranked Robin elect different winners, deliberately spread across candidate-field size, electorate size, and structure — with RCV-IRV, Approval and Plurality on the same ballots. Every winner is the LH engine's (from each case's _tabulated mirror). Each YAML's scenario_description states the exact cause of its divergence. Empirical companion to the STAR-vs-RR simulation.

Regenerating this folder? Take the winners from the engine — star_winner_engine() in the simulation, or each case's _tabulated mirror. The simulation's fast numpy model (star_winner()) now implements the engine's tie-breaks and is test-checked against it, but the engine is right by construction and cannot drift; an earlier version of that model did drift, and mislabelled one of these 30 at birth. Then verify with:

python STARVote_LH_tabulation_engine/tools_adam/scripts/check_star_vs_rr_labels.py
python STARVote_LH_tabulation_engine/tools_adam/scripts/check_star_vs_rr_causes.py

The first checks all five places a sample names a winner — expected_winners, the title, the description, the _tabulated mirror and the table below — against a real tabulation (--fix relabels the yamls). The second checks what the descriptions claim about why the methods diverge: that each cycle chain (A>I>G>A) is a real chain of pairwise wins that closes, that a cycle sample really has no Condorcet winner, and that a dark horse really is the Condorcet winner with the score rank, totals and missed finalists it claims. A sample can name both winners correctly and still assert a link that does not exist, so the two are complementary — winners vs. causes. tests/test_star_vs_rr_labels.py and tests/test_star_vs_rr_causes.py run them on every commit.

The spread (what varies)

  • Candidate field: 5 × 3, 6 × 5, 9 × 7, 10 × 10 (few → large).
  • Electorate: few (~15–30), medium (~45–150), and large (up to ~600) voters.
  • Structure: 15 grouped (a few voter factions/blocs) vs 15 ungrouped (independent random ballots).
  • Flavor: 21 cycles + 9 dark horses — dark horses are rarer and cluster at higher candidate counts and grouped electorates (a broadly-liked compromise only emerges when there's structure to compromise around).

The two causes (also spelled out per file)

  • Cycle — no Condorcet winner (rock-paper-scissors). RR resolves by Copeland/margin; STAR by score-runoff. A structural coin-flip electorate.
  • Dark horse — a Condorcet winner exists but misses STAR's score top-two: a broadly-preferred, low-intensity compromise. The preference-vs-support split; a real RR limit.

Who sides with whom

On these 30 STAR≠RR elections, the other three methods scatter — there is no clean alignment:

method agrees w/ STAR agrees w/ RR picks a third candidate
Approval 12 8 10
RCV-IRV 12 8 10
Plurality 13 8 9

The honest read: all three lean mildly toward STAR's winner over RR's, but the striking fact is that about a third of the time each of them elects a candidate that is neither STAR's nor RR's winner — the field fragments, especially with more candidates. (In a narrower 3–5-candidate sample the split looks cleaner — Approval→STAR, IRV/Plurality→RR, matching the support-vs-order intuition — but that alignment washes out once large fields are included. So the robust claim is only the modest one: STAR and RR genuinely disagree here, and the other methods don't reliably break the tie for either side.)

The table

flavor cands voters electorate STAR RR IRV Appr Plur file
cycle 3 15 random A B A A A cycle_C03_fewV15_noise_1
cycle 3 15 random A C A A A cycle_C03_fewV15_noise_2
cycle 3 45 random A B A B A cycle_C03_medV45_noise_1
cycle 3 45 random B C A C A cycle_C03_medV45_noise_2
cycle 5 15 random B E A E B cycle_C05_fewV15_noise_1
cycle 5 15 random A E A D A cycle_C05_fewV15_noise_2
cycle 5 28 grouped A C A A A cycle_C05_fewV28_bloc_1
cycle 5 45 random A B A E A cycle_C05_medV45_noise_1
cycle 5 45 random D B B B A cycle_C05_medV45_noise_2
cycle 7 15 random D A A A A cycle_C07_fewV15_noise_1
cycle 7 28 grouped D A D A D cycle_C07_fewV28_bloc_2
cycle 7 149 grouped F C F C A cycle_C07_medV149_bloc_2
cycle 7 598 grouped C E F A A cycle_C07_largeV598_bloc_1
cycle 10 15 random A C H A A cycle_C10_fewV15_noise_1
cycle 10 15 random J F B F G cycle_C10_fewV15_noise_2
cycle 10 28 grouped C F C C C cycle_C10_fewV28_bloc_1
cycle 10 29 grouped C B B A B cycle_C10_fewV29_bloc_2
cycle 10 45 random E G A C A cycle_C10_medV45_noise_1
cycle 10 45 random A I A A A cycle_C10_medV45_noise_2
cycle 10 148 grouped E G J B J cycle_C10_medV148_bloc_1
cycle 10 149 grouped I H A A C cycle_C10_medV149_bloc_2
dark horse 3 15 random A C A A A darkhorse_C03_fewV15_noise_1
dark horse 5 599 grouped A E A A A darkhorse_C05_largeV599_bloc_1
dark horse 7 30 grouped D C C D C darkhorse_C07_fewV30_bloc_1
dark horse 7 45 random E A A E A darkhorse_C07_medV45_noise_1
dark horse 7 147 grouped F D D F D darkhorse_C07_medV147_bloc_1
dark horse 7 597 grouped D E B C B darkhorse_C07_largeV597_bloc_1
dark horse 7 598 grouped E G G A G darkhorse_C07_largeV598_bloc_2
dark horse 10 598 grouped G F F D F darkhorse_C10_largeV598_bloc_1
dark horse 10 599 grouped B E C B E darkhorse_C10_largeV599_bloc_2

Caveats (read before quoting)

  • Auto-generated stress tests, not real elections — they show the mechanism; for a rate, use the simulation.
  • The IRV column is noisy. On small/tied score ballots the rank IRV reads is often decided by candidate-priority tie-breaking (the engine flags it per case). STAR, RR and Approval read scores/pairwise directly and are robust.
  • RR = Copeland (LH's margin-then-lot tiebreak); STAR/RR winners are the engine's. Sincere normalized 0–5 scores.
  • Each case is a normal STAR YAML — re-run it with the LH engine to see the full [Divergence from STAR] block, the pairwise matrix, and the IRV rounds.