Margin of error is per-candidate, the sample matters more than the size, and 'likely voters' is a model, not a fact.
A poll says 48–45 with a margin of error of ±3. Sounds like a 3-point lead with a 3-point cushion — basically safe, right? Wrong twice. The ±3 applies to EACH number separately: the 48 could really be 45, and the 45 could really be 48. The gap between two candidates can swing roughly double the stated margin. A 3-point lead inside ±3 is a coin flip wearing a suit.
Rule: when the lead is smaller than twice the margin of error, the honest headline is 'too close to call' — which is exactly why you'll rarely see that headline.
A sloppy sample of 10,000 loses to a careful sample of 800. What matters is whether the people polled look like the people who will actually show up. Check three things in the fine print: registered voters or likely voters or all adults (these are different universes); how it was conducted (live calls, online panel, text); and who paid for it.
'Likely voters' deserves special suspicion: it's not a fact about the world, it's the pollster's MODEL of who turns up. Two pollsters with identical raw data and different turnout models will publish different races.
Polls are weather readings, not prophecies. Any single poll — especially a surprising one — is most useful as a data point inside an average. The surprising poll gets the headlines precisely because it's surprising, which is to say: unrepresentative, until corroborated.
This is the wire's corroboration rule wearing a different hat: one source is a rumor; one poll is a vibe.
Next poll you see, find the margin of error and double it. If the lead fits inside, say out loud: 'coin flip.' Notice how different that feels from the headline.
Share this lesson freely — the method only works if it spreads. Watch it in action on the live Wire, where every rank shows this exact work.