Skip to content

Reading paths for reviewers

Each path is four pages, in the order that answers the usual questions first.

Data science

How signals, models and evaluation work, and what a score does and does not claim.

  1. The consumer graph
  2. How models work
  3. Evaluation and release
  4. Held-back groups

Security and privacy

What enters, what leaves, how clients are separated and who can see what.

  1. What leaves the platform
  2. Architecture and isolation
  3. Access and roles
  4. Privacy and permitted use

Scientific audit

The research each claim rests on, how strong it is, and what Moonbrush has measured itself.

  1. Research overview
  2. Evaluating ranking models
  3. Measuring what a campaign caused
  4. Claims register

Marketing operations

Files in, packages out, and how results come back and get counted.

  1. Matching
  2. Package and file formats
  3. Campaigns and packages
  4. How a result is counted
Question Short answer Read
Do you resolve people probabilistically? No. Resolution is deterministic, and a detail matching several people binds to no one. Identity resolution
Is a score a probability? No. Catalog scores are rankings and bands from 0 to 1, dated and versioned. Reading a score
How are models tested? Each catalog release scores a sample first and must pass evidence, size, overlap, determinism and language gates. Models are evaluated on a held-out test set. Evaluation and release
Does my data train models for other clients? No. Your results are used only for you and never change the shared catalog. Outcome models
What leaves the platform? Campaign packages carry SHA-256 hashed identifiers. Workspace exports are a separate, permissioned path. What leaves the platform
How are clients separated? Every query is scoped to the caller’s organization and fails safe. Each organization has its own warehouse identity, and cross-organization reads are denied by platform permissions. Architecture and isolation
Do you report lift? No. Reports describe response among the people reached and never say a campaign caused a result. Held-back groups
What is the scientific basis for your models? Published research on behavior, habit, record linkage, evaluation and experiments, graded and cited in seven research notes, with a register mapping each claim to its evidence. Research
  • The SOC 2 Type I report. A SOC 2 Type II examination is underway.
  • Completed security questionnaires.
  • A technical session on model design, evaluation records and the release process.
  • The study design and acceptance record behind the published retrain result.

Write to security@moonbrush.com for security material, or hello@moonbrush.com for anything else.

Every figure on these pages carries the date it was measured. See Published figures.

© 2026 Moonbrush Inc. Moonbrush, atlas, stage and Gravity are pending trademarks of Moonbrush Inc.