Claims register

The register lists the claims in these docs that rest on science rather than on a description of the product. For each, it gives the research behind it, the grade of that research and the evidence Moonbrush itself holds. Where a row pairs a research finding with a product rule, the grade applies to the research finding only; the product rule is described, not graded. Grades are defined on the research overview.

The last column uses three entries. Published means a dated figure or a product rule is in these docs. Under NDA means the evidence exists and is shown to qualified reviewers in a technical session. Not published means these docs make no claim that Moonbrush has measured it.

Behavior and models

Table 1. Claims about behavior and what models read.

Claim	Where it is made	Research basis	Grade	Moonbrush evidence
Stated intentions predict behavior, substantially but imperfectly, and on average	Note 1, Reading a score	(Armitage and Conner, 2001; Sheeran, 2002; Webb and Sheeran, 2006)	Established	Product rule: scores are ranks and audience groups are the reliable unit
Search and fine-grained behavioral data forecast outcomes	Signals and intent topics, note 1	(Goel et al., 2010; Moe, 2003; Martens et al., 2016)	Supported	Published: one retrain result (note 6). Under NDA: evaluation records per model
Repeated behavior in a stable context predicts future behavior	The model catalog, Signals and intent topics, note 2	(Ouellette and Wood, 1998; Wood et al., 2002)	Established	Not published: no habit model’s evaluation is published
Life changes and decision stage open windows in which response differs	The model catalog, note 2	(Andreasen, 1984; Verplanken and Roy, 2016; Lambrecht and Tucker, 2013; Sahni et al., 2019)	Supported	Not published: window spans are not published and no Moonbrush timing study is claimed
Why people buy has a structure that recurs across categories	The model catalog, note 3	(Gutman, 1982; Reynolds and Gutman, 1988)	Reported	Not published: the evidence gate checks that a reason’s behavior is witnessed in the data, which does not validate the model
Stable dispositions can be estimated from behavioral records	The model catalog, note 3	(Youyou et al., 2015; Kosinski et al., 2013)	Reported	Not published: the evidence gate checks that a trait’s behavior is witnessed in the data, which does not validate the model
Messages matched to a person’s characteristics persuade more	Creative plan, note 3	(Hirsh et al., 2012; Teeny et al., 2021) in a rating study; (Matz et al., 2017) in the field, challenged by (Eckles et al., 2018)	Contested in the field	Not published: no persuasion effect of matching is claimed

Audiences, identity and privacy

Table 2. Claims about how audiences are formed and how people are identified and protected.

Claim	Where it is made	Research basis	Grade	Moonbrush evidence
A useful segment is identifiable, substantial, accessible, stable, responsive and actionable	Audience groups, note 4	(Smith, 1956; Wedel and Kamakura, 2000)	Supported (standard text)	Published: group card fields and product rules
Rates on small groups swing widely by chance	Audience groups, Reporting principles, note 4	(Wainer, 2007; Tversky and Kahneman, 1971)	Established (sampling theory)	Published: the ratio-and-gap rule, the rule that ratios need a base, and a seeded simulation. Under NDA: thresholds
Data-driven decisions can have a disparate impact on protected groups, even without intent	Audience groups, Privacy and permitted use, notes 3 and 4	(Barocas and Selbst, 2016; Ali et al., 2019; Speicher et al., 2018)	Supported	Published: demographics describe but are not evidence (since August 2, 2026); Data Use Policy terms on permitted use and sensitive fields
Deterministic matching trades missed links for fewer false links	Identity resolution, note 5	(Fellegi and Sunter, 1969; Christen, 2012)	Established	Published: deterministic rules and the match rate shown before a paid build
A hashed email or phone number is a pseudonym, not anonymous	What leaves the platform, note 5	(Demir et al., 2018)	Established	Published: hashing rules, export-scoped row keys, suppression on every path
Combinations of ordinary facts can re-identify people	Reporting principles, note 5	(Sweeney, 2000; Sweeney, 2002; Narayanan and Shmatikov, 2008)	Supported	Published: small counts are not printed exactly. Under NDA: the threshold

Evaluation and measurement

Table 3. Claims about how models are graded and how results are read.

Claim	Where it is made	Research basis	Grade	Moonbrush evidence
A predictive model must be graded on cases it has not seen	Evaluation and release, note 6	(Shmueli, 2010; Hastie et al., 2009)	Established	Published: models are evaluated on a held-out test set. Under NDA: the study design
AUC is the probability of ordering a random pair correctly	Evaluation and release, note 6	(Hanley and McNeil, 1982; Fawcett, 2006)	Established (mathematics)	Published: AUC 0.87 from one full retrain
Top-decile lift reads what a ranked list holds where it is used	Evaluation and release, note 6	(Ling and Li, 1998; Neslin et al., 2006)	Supported	Published: 5.04× from one full retrain
A score that ranks well is not thereby a calibrated probability	Reading a score, note 6	(Niculescu-Mizil and Caruana, 2005)	Established (mathematics)	Published: scores are ranks; no probability claim
The people a model ranks highest are not necessarily the people a campaign changes most	Notes 6 and 7	(Ascarza, 2018; Gutierrez and Gérardy, 2017)	Supported	Published: ranking results and campaign results are kept apart
Response among the people reached is not the effect of a campaign	Held-back groups, note 7	(Holland, 1986; Rubin, 1974)	Established	Published: reports print response, not lift
Giving the whole result to the last touch is a counting rule, not an effect	How a result is counted, note 7	(Holland, 1986; Berman, 2018)	Established (definition)	Published: “did it from our ads” is defined as a last-arrival count
Observational ad measurement can be badly biased; randomized comparison is the standard	Held-back groups, note 7	(Gordon et al., 2019; Blake et al., 2015; Lewis and Rao, 2015)	Established	Published: the held-back group design. Not published: no causal estimate
Ad effects are small relative to noise, so comparisons need very large groups	Held-back groups, note 7	(Lewis and Rao, 2015; Kohavi et al., 2009)	Established	Published: computed detectable differences (note 7, figure 2)
Fixing the rules before seeing the data guards against results that follow the data	How a result is counted, note 7	(Nosek et al., 2018)	Supported	Published: each Outcome fixes counting, attribution and target before launch (the principle, not a preregistered study)
Documenting each model makes results traceable	How models work, note 6	(Mitchell et al., 2019; Gebru et al., 2021)	Supported as practice	Published: a card for every model; versions recorded on every score and package

Figures in the register

The dated figures behind these claims are defined in Published figures. The retrain result is reported as published: Moonbrush-reported result from one retrain cycle. The study design and acceptance record are available under NDA.