Every published poll carries one margin of error, quoted for the biggest number on it. That figure is then used to read a table where most of the numbers are small, several of the differences matter more than the levels, and one party is usually sitting next to a cliff.
Three things go wrong, and they compound:
The margin depends on the number. At n = 1000 a party on 50% has ±3.1 points; a party on 5% has ±1.4. In points the small party looks more certain. As a share of itself it is five times less.
A lead is less certain than either number in it. Two shares in one poll are negatively correlated — when one is over-sampled the other is under-sampled — so subtracting them adds to the variance. “35 to 33, a two-point lead” carries a margin of ±5 points.
The answer is seats, not percentages. And seats come off a cliff at 5%. A party on 4.8 and a party on 5.2 are one margin of error apart and about eight seats apart.
So this page does not quote a margin. It draws the sample again four thousand times, runs the real Sainte-Laguë allocation on every draw, and reports what the seats did — in the unit the question was asked in.
The example parties are called A, B and C on purpose. Putting an invented poll result next to a real party's name makes a thing that can be screenshotted and passed on as if it were real, and the mechanics do not need it: paste today's actual numbers into the box and they work the same. Nothing is uploaded — this page has no server.
The poll
What it does not know
Sampling error is not the main error in a poll. Who answers the phone, who refuses, who is weighted up to stand for the people who refused, and who actually turns out — every one of those is usually a bigger source of error than the sampling, and none of them is on this page. Everything here is therefore a floor on the uncertainty, not a measure of it.
It assumes a simple random sample. Real polls are stratified, quota'd and weighted, which typically makes the effective sample smaller than the number of people spoken to. Where a pollster publishes a design effect, dividing the sample size by it before typing it in gets closer to the truth.
It knows nothing about electorates. Electorate seats are typed in, not predicted, and every one of them is its own contest with its own uncertainty — including the one that decides whether a small party gets in through the second door at all.
The tie-break is not the legislated one. Two parties with exactly the same quotient are separated here by the larger vote and then by name. The Electoral Act settles it by lot. It has never mattered and it would take a coincidence to the fourth decimal place, but it is not the rule.
A coalition is not arithmetic. Reaching a majority on paper is a necessary condition and nothing else. Who will work with whom, and on what terms, is not a number and is not here.