Quality, coverage and freshness
Data reaches a client only after it passes checks that were written down before the run. When a check fails, the previous version keeps serving.
Catalog releases
Section titled “Catalog releases”A new version of the Gravity catalog goes through these steps.
- Score a sample first. The new version scores a sample of the population before anything else.
- Grade against acceptance bands. Each model’s measured size on the sample is graded against the band set for it in advance.
- Run the population. The full run is checked against the sample, so a difference in scale or shape is caught.
- Stage, gate, swap. The release is staged, must pass every gate, then swaps in. Each surface can roll back on its own.
Every catalog model also passes gates on its own before it can be picked:
| Gate | What it checks |
|---|---|
| Evidence | The behavior the model claims is actually witnessed in the data. |
| Size | The model finds a usable number of people, inside its expected range. |
| Overlap | The model is not a near-copy of another model. |
| Determinism | The same inputs give the same scores. |
| Language | The model’s name and card describe what it measures and nothing more. |
Monthly graph cycles
Section titled “Monthly graph cycles”Each monthly graph delivery is rehearsed on a sample, must produce zero duplicate people, and is refused if it would drop an unusual share of known people. See The consumer graph.
Coverage
Section titled “Coverage”Coverage differs by model, because a behavioral model can only score people whose behavior is visible to it. Each model card reports, in plain words, how the model reads today:
| Reading | Meaning |
|---|---|
| Works | Evidence and size are inside the expected range. |
| Evidence short | The behavior is witnessed, but thinly. |
| Not read yet | The model has not been read in this release. |
| Under the minimum | It finds too few people to use in a campaign. |
| Nobody | It finds no one in this release. |
A model below the minimum, or without witnessed evidence, is not offered, and its card says why.
Freshness
Section titled “Freshness”- The behavioral feed is read daily, about a day behind.
- Each model card states how often that model is re-scored.
- Every score carries its model version and date, so a delivered file always says when its scores were made.
- Buying windows expire on their own, so a person who has stopped researching leaves the window without anyone removing them.