Method
How this is measured, and what it cannot tell you.
Published before the first figure, so that no rule here can be chosen after seeing a number it would flatter.
The panel
Panellists register with an email address and a mobile number, both verified. Each person records at most one vote per election per day, and the answer locks on save. A missed day is recorded as a gap; we never carry an answer forward.
Weighting
An app that asks people to vote on politics every day recruits the politically engaged and the well educated. That is not a small skew, and it is the failure mode the 2019 Australian polling inquiry identified. We collect education and past-vote recall because they are the two variables that most improve a weighted estimate, and we publish weighted and unweighted figures together so the correction is visible rather than implied.
Location
Location is captured at registration and at each vote, and it never blocks anyone from voting. Readings resolve to three states: in jurisdiction, out of jurisdiction, and unknown. Unknown is reported as its own category and is never folded into "out" — a fifth of any panel will sit there because they declined the prompt, and treating that as evidence of anything would be wrong.
We do not keep coordinates. A reading is resolved to a region in memory and the coordinates are discarded. Location is also weak evidence: it is straightforward to spoof, so we treat it as a disclosure statistic and as one input among several, not as proof of eligibility.
What we publish
Every reading appears twice — raw and matched — because the difference between them is the part no single-night poll can show you. Cells below our suppression floor are shown as a dash rather than a number, since a small enough crosstab describes an identifiable person.
Privacy, stated accurately
No one can see how an individual panellist voted in anything we publish. That is a different claim from "we do not know", and we will not make the second one: the daily record has to be linked to a person or none of the matched-panel analysis on this site would be possible.