Research

Moonbrush makes claims that rest on science: that behavior before an action says something about the action, that a model ranks people well, that a result was counted correctly and that an identifier is protected. This section sets out the published research each claim rests on, how strong that research is, what Moonbrush has measured itself, and where the support stops.

It is written for auditors. Every source is cited in full, with 84 works in the bibliography. Every computed figure states its formula and parameters, every simulation states its seed, and every limit is stated next to the claim it limits.

The notes

Each note is formatted as a short paper. Every note has an abstract, the research, how it bears on the product, limits, the questions an auditor should ask, and references; most also have figures and a table of what the evidence does and does not support.

Table 1. The research notes and the question each answers.

No.	Note	The question it answers
1	Behavior before the act	Does what people research and browse carry information about what they later do?
2	Habits and buying windows	Why read repeated behavior and recent episodes of research differently?
3	Traits, reasons and message fit	Is there structure to why people buy, and does matching a message to it work?
4	Splitting an audience	How should an audience be divided, and when is a difference between groups real?
5	Identity and hashing	Why deterministic matching, and what does hashing protect?
6	Evaluating ranking models	What do an AUC and a top-tenth rate mean, and what can make them wrong?
7	Measuring what a campaign caused	Why is response not effect, and what would a credible comparison need?

Kinds of evidence, kept apart

Table 2. The kinds of evidence these notes use, and how each is marked.

Kind	What it is	How it is marked
Published research	Peer-reviewed studies, standards and standard texts	An author-year citation linked to the bibliography
Moonbrush figures	Dated measurements from the product, such as the size of the consumer graph or one retrain’s AUC	“Moonbrush published figure” under the chart; definitions in Published figures
Values from studies	Numbers reported in the cited studies, charted as reported	“Values reported in the cited studies” under the figure, with the citations in the caption
Illustrations	Curves computed from a stated formula, seeded simulations and schematic drawings	“Computed”, “Simulated” or “Schematic” under the figure, with the formula, seed or a note that the shapes are invented

Evidence grades

Where a note grades the research behind a claim, it uses four grades.

Table 3. Evidence grades.

Grade	Meaning
Established	Supported by meta-analyses, by several large randomized studies, or by mathematics
Supported	Replicated across independent studies, mostly observational or in the laboratory, without a substantial published challenge
Reported	Reported in peer-reviewed work, without the independent replication the grades above require
Contested	Reported, but with a published methodological critique or conflicting results

A grade describes the research, not Moonbrush. Whether a particular Moonbrush model captures an established effect is a separate question, answered by its evaluation. See the claims register for each claim, its grade and Moonbrush’s own evidence.

What this section does not do

• It does not claim the literature proves Moonbrush’s models work. Research establishes premises. Evidence that a specific model works is its evaluation record, published where it can be and otherwise available to qualified reviewers under NDA.
• It does not publish model internals. Feature construction, model types, training data composition, window lengths and thresholds are not published here. Where a note shows a formula for a Moonbrush concept, it is the definition of the concept, not the implementation.
• It does not hide disagreement. Where a finding is contested, the note says so and cites the critique, as note 3 does for field studies of message matching (Eckles et al., 2018).

Versions and corrections

Each note carries a version and the date it was checked against the product. A change to a note is recorded in the changelog. If you find an error in a citation, a figure or a claim, tell us and we will correct it and record the change.

Claims registerEvery scientific claim in these docs, its basis, its grade and Moonbrush's own evidence.

ReferencesThe full bibliography, with DOIs and the pages that cite each work.