On result day, every channel has a number. Some of those numbers were built; most were guessed. The difference is invisible on television and unmistakable in the method - which is why we talk less about our projections and more about how they are made.
An exit poll is not a prediction machine. It is a measurement of one thing: what voters who just voted say they did, collected in a way that mirrors the constituency itself. Done honestly, it produces a projection with a known confidence range. Done casually, it produces a headline that ages badly overnight.
The sample is designed before polling day, not on it
Our booth selection is settled well in advance - balanced by region, community, urban and rural mix, and past voting behaviour. On the day itself there is no improvisation: every interviewer knows their booths, their quota, and their script. If a booth cannot be covered, its designated substitute is already on the sheet.
People do not always say what they did
The best-known failure mode of exit polling is the polite answer. Voters may name the locally dominant party, the safer choice, or simply decline. We counter this the unglamorous way: anonymous response slips instead of spoken answers where needed, interviewer matching, and weighting rules built from a decade of comparing our field responses against declared results.
"Anyone can publish a number; standing behind it the next morning is a different matter."
Exit poll desk, Political Laboratory
Why the read sharpens through the day
Responses stream into a single dataset from the first hour of polling. Early numbers are noisy and treated that way. By mid-afternoon, patterns stabilise; by the close of polls the projection carries its final weighting and its confidence range. Clients receive not just a seat count but the reasoning - which booths moved, which communities shifted, and how sure we are.
That is what a 95% accuracy record actually is: not luck repeated, but a method repeated. The number is only the last line of a long, careful day.
