Reproducing the numbers¶
star-abstention-probe.ts runs BetterVoting's real Star() tabulator over four ballot sets and prints the tally counts, scores, pairwise counts, winner, tie-break type, and the runoff/score percentages that the results page would derive from them.
Every number in 03-reporting-anomalies.md and the invariance claim in 02-blast-radius.md comes from this, run against Equal-Vote/bettervoting at 8d2b3f9.
Setup¶
git clone --depth 1 https://github.com/Equal-Vote/bettervoting.git bv
mkdir probe && cd probe && npm init -y && npm install tsx fraction.js
mkdir -p node_modules/@equal-vote
ln -s ../../../bv/packages/shared/src node_modules/@equal-vote/star-vote-shared
ln -s ../probe/node_modules ../bv/node_modules
cp /path/to/star-abstention-probe.ts run.ts
The tabulators only need fraction.js at runtime — everything imported from @equal-vote/star-vote-shared is type-only and erased, so no build of the shared package is required.
The three runs¶
# 1. Today's behaviour, unmodified
npx tsx run.ts
# 2. Edit (a) — all-equal is a cast vote. Patch Star.ts:13 makeAbstentionTest(true) -> makeAbstentionTest()
npx tsx run.ts
For run 3, replace the body of makeAbstentionTest in packages/backend/src/Tabulators/Util.ts with the "only truly blank abstains" rule:
const raw = Object.values(vote.marks);
return raw.every(m => m == null);
# 3. Edit (b) — only a truly blank ballot abstains
npx tsx run.ts
Results¶
| Ballot set | Today | Edit (a) | Edit (b) | |
|---|---|---|---|---|
EX1 3×5,3 2×3,5 4×5,5 |
tally / abst | 5 / 4 | 9 / 0 | 9 / 0 |
| scores | A 21, B 19 | A 41, B 39 | A 41, B 39 | |
| runoff | A 60% B 40% Eq 0% | A 33.3% B 22.2% Eq 44.4% | same | |
| winner | A (no tiebreak) | A (no tiebreak) | A (no tiebreak) | |
EX2 3×5,3 2×3,5 4×0,0 |
tally / abst | 5 / 4 | 5 / 4 — unchanged | 9 / 0 |
| score % | A 84% B 76% | A 84% B 76% | A 46.7% B 42.2% | |
| winner | A | A | A | |
EX3 all-flat 1,1,1…5,5,5 |
tally / abst | 0 / 5 | 5 / 0 | 5 / 0 |
| scores | all 0 | all 15 | all 15 | |
| runoff | — (page says no votes cast) | 0%, 0%, Equal 100% | same | |
| winner | Anchovy (random) | Anchovy (random) | Anchovy (random) | |
EX4 3×5,3 2×3,5 2×0,null 2×null,null |
tally / abst | 5 / 4 | 5 / 4 — unchanged | 7 / 2 |
| score % | A 84% B 76% | A 84% B 76% | A 60% B 54.3% | |
| winner | A | A | A |
Two things fall straight out of the table:
- The winner never changes — in any ballot set, under any of the three rules, including the all-flat election that resolves by random tiebreak.
- Edit (a) and edit (b) affect disjoint ballot sets. EX1 and EX3 move under (a); EX2 and EX4 move only under (b). They are genuinely two changes.
Caveat¶
This exercises the backend tabulator only. The runoff and score percentages in the table are computed by applying ResultsBarChart's own formulas (percentDenominator = sum of all bars; majority marker = sum excluding the last bar, halved) to the tabulator's real output. The NaN% prediction for the EX3 pie chart is read from ResultsPieChart.tsx and has not been rendered in a browser here — that check still wants doing on a branch.
nan-fix-verify.ts — the #1035 before/after¶
Added 2026-08-19, for the zero-denominator fix on branch fix/1035-runoff-zero-denominator (47d241a4). It runs the real Star() over two degenerate ballot sets and two controls, then prints, per set, what each runoff render path produced before the fix and after it — the % Between Finalists column, its total row, and whether the pie chart has anything to draw.
The point that closes the caveat above: it does not transcribe formatPercent. It reads packages/frontend/src/components/util.tsx at run time, extracts the function's source, erases its TypeScript annotations and evaluates it, so the "after" column is the shipped function and cannot drift from the file. The two remaining expressions — the total-row cell and the pie's isEmpty — live inside JSX and are transcribed verbatim, one line each.
# from a bettervoting checkout with node_modules already installed
cp nan-fix-verify.ts <checkout>/nan-fix-verify.ts
cd <checkout> && npx tsx nan-fix-verify.ts
Full output and what it proves: ../../../issues/1035-runoff-zero-denominator-fix.md.
The browser check is still outstanding — the same gap this README's caveat named in August.