A level and a contrast
A level is a share standing on its own: how much of a place would say yes. A contrast is a difference inside one reading: option B set against option A, renters set against homeowners, the same question with and without a piece of news.
Where a model’s error sits
A language model reading a character sheet can lean the same way for everyone on the panel — a little more agreeable, a little more taken with something new. A lean like that moves a level directly, and nothing in the reading can see it: the interval prices how much information the reading holds, not whether the model read the place well.
In a contrast, both sides are read by the same model, from the same kind of brief, often for the same twins. A lean they share lands on both sides, so it does less to the difference between them than to either level. The contrast is the part of a result that a shared lean has the least room to fake.
Same twins on both sides
Where it can, a reading puts both sides of a comparison to the same twins. A split puts two versions of a message to the same panel, so the gap between them is not a gap between two different samples. A what-if puts the same question to the same population with and without the news. Options in one question are weighed by each twin side by side.
When a contrast is too close to call
A contrast is only worth reporting if the reading can separate the two sides. When two options or two groups sit inside each other’s intervals, the reading says it cannot tell them apart rather than naming one ahead. Every result also carries a read-quality grade — Tight read, Comparable, Directional only or Incomplete — that says how much weight it can bear.
How to use a reading
Use a reading to compare options, to find which groups differ and in which direction, and to see which way a piece of news pushes. Do not use it as a forecast of sales, turnout, demand or an election result, and do not quote a reading’s share as if it were the share of a real place.