How models work

Gravity is the model catalog behind atlas and stage. A model is a versioned, dated ranking from 0 to 1 over the people it can see, read from the signal registry.

Kinds of model

Kind	What it reads	Example
Buying window	A recent, still-warm episode of research that expires on its own	Is about to install a whole-home system
Habit and shopping manner	Repeated behavior over weeks, relative to everyone else	Waits for the markdown window
Trait	A stable lean between two ends of a scale	Named in the product
Reason	Why a person buys, in any category	Named in the product
Work	What a person does for work and how	Named in the product

See The model catalog for each kind.

Scores and picks

Each model scores every person it can see. Scores are rankings and bands, not probabilities: a higher score means the person ranks higher on the model, not that they have a stated chance of acting.

A model is used as a pick: the band of people at or above a measured cut. Picks are what search filters on and what audience groups are described with. See Reading a score.

A person a model cannot see gets no score. Unknown is never shown as low.

Model cards

Every model carries a card. The card is the contract for what the model means. The practice matches model cards as proposed for reporting machine learning models (Mitchell et al., 2019).

• Definition. What the model finds, in one plain sentence.
• What it means and does not mean. The limits of the claim, stated next to the claim.
• Evidence. The behavior that supports it, and the cue terms behind it.
• Size. Expected and measured size in the current release.
• Window and cadence. The span of behavior it reads and how often it is re-scored.
• Reading. Works, evidence short, not read yet, under the minimum, or nobody. See Quality, coverage and freshness.
• Why it is withheld. When a model is not offered, the card says why.

Catalog models and Outcome models

Property	Catalog models	Outcome models
Learns from	Population behavior	One client’s campaign inquiries for one Outcome
Who sees them	Every client	Only that client
Changed by client results	Never	Only that client’s own, after a person approves training
Output	Scores and picks	A starting audience in Search: people most like the ones who inquired
Status	Live	Being built

See Outcome models and Evaluation and release.

Versioned and dated

Every score carries its model version and the date it was made. A delivered package records which version each person was sent under, so a result can always be traced to the exact scores that produced the audience.