Scores you can
actually compare.
A 0.72 means the same strength of evidence for a truck buyer, a cruise enthusiast, or a registered voter. One ruler across tens of thousands of attributes — and audience logic becomes plain arithmetic.
Trained models emit probabilities calibrated to each target's base rate. A 0.85 from a truck-buyer model and a 0.85 from a cruise model do not mean the same thing, so audiences cannot be defined as arithmetic across attributes. This paper shows how a single closed-form measure of association, mapped onto one normalized scale with a fixed neutral point at 0.25, makes every score commensurable — and why that property is the foundation of the super audience.
Why two model probabilities aren't comparable.
Every separately trained classifier learns to its own base rate. A rare attribute (owning a Ferrari) and a common one (owning a phone) produce probabilities that live on different scales. Post-hoc calibration — Platt scaling, isotonic regression — squeezes them onto a probability axis, but the axis is still per-model. A marketer cannot write a rule like "truck ≥ 0.72 AND outdoors ≥ 0.68" and trust that the two thresholds carry the same meaning.
A fixed neutral point, everywhere.
The Brain measures association the same way for every attribute and maps it through one normalization. The result is a universal 0-to-1 scale with a fixed meaning: 0.25 is "no signal," 1.0 is "known true," 0.0 is "strong negative." Because zero association lands at exactly 0.25 for every attribute, a score's distance above 0.25 is directly comparable across all tens of thousands of attributes.
Audience logic becomes legible.
Once scores are commensurable, an audience is just arithmetic on the scale — and the same arithmetic works across every attribute. "Truck above 0.72, outdoors above 0.68, recent purchase below 0.25" is a precise, comparable definition, not a hope. This is what makes the super audience possible: stacking behavior, motivation, channel, and creative on a single, comparable ruler.
Comparability isn't a convenience. It's the precondition for treating a person — not just a segment — as the unit of targeting, because every person carries all attributes on the same ruler at once.
Stop comparing apples to each model's calibration.
The base-rate calibration tax is a structural cost of training one model per attribute. A single measured scale dissolves it — and in doing so, unlocks the rest of the series: explainability, super audiences, and stable, comparable campaign measurement.