Gravity · the model engine

Probabilities with reasons, and a label that says where they’re weak.

Gravity is the model engine under atlas and stage. It estimates who is likely to act, why, how they decide, when they can hear it and whether reaching them will change anything, then proves each model on outcomes you own before anyone uses it.

100: models in the live catalog
5: model families
17,000+: named signals in one registry
7,015: active consumer topics, 25 categories
62.6B: resolved behavioral events, refreshed daily
8,087: client outcome labels in the closed loop

Counts as of the August 19, 2026 audit and the October 2026 capabilities review.

The registry

17,000+ signals. Each one defined, checked and allowed.

Every signal in the registry has a written definition, a fill-rate check and a reason to exist. Models draw from the same registry, so a score can always be traced back to the evidence behind it.

Behavioral events
62.6B resolved events, refreshed daily: what people read, research, compare and come back to, and when.
Intent topics
7,015 active consumer topics in 25 categories, from a dictionary of 10,636.
Purchase cadence
How often, in what volume and in which categories people buy, not just whether they did.
Roles and employers
Professional titles and changes in them, at the person level.
Brand affinities
Which brands and tools people engage with, and which they leave.
Your outcomes
Conversions you agree to share. 8,087 client outcome labels return to the loop today.

Never in the registry

Modeled ethnicity or political ideology, as an input or a selection
Data about minors
Protected health information. No PHI is ingested, by design
Identifiable client data in a model other clients use

Gravity

The model family underneath both.

100 models in the live catalog, in five families, all drawing on one registry of 17,000+ signals. Each score is a calibrated probability from 0 to 1. They are probabilities, not verdicts, and every one carries a label that says where it is weak.

Family What it estimates What you do with it

Propensity How likely a person is to take a specific action within a window Who to reach first

Motivation The reasons that may matter most to them right now Which message to lead with

Decision style How they approach a choice: research depth, pace, who they consult How much to say, and how fast

Receptivity Openness to a message or a moment, and when they can hear it When to reach them

Persuadability Potential to change behavior, rather than act as they already would Where spend moves the number

No model ships until it beats your baselines.

Judged on outcomes you own, against two baselines and your current approach, then watched every day. The validation report covers ROC and AUC, precision and recall, lift by decile, calibration, stability across rebuilds and the business outcome on your holdout.

0.87AUC, held-out ranking performance of the published model

5.04×lift in the highest-scoring 10% against the population rate

From one full higher-education retrain cycle. Moonbrush-reported results from one full retrain cycle. Study details and the acceptance record are available under NDA.

01 Your holdout Outcome labels you own, held out and never trained on
02 Candidate model Gravity scores every record in the holdout
03 Three comparisons Against a demographic baseline, a lookalike baseline and your current model: AUC, lift and calibration
04 Publish or kill It ships versioned, with a label that says what it is, what it isn’t and where it is weak. Or it is killed and the reason is logged.
05 Monitor daily New outcomes recalibrate the model. Drift triggers a fresh review.

Persuadability

The most expensive person to reach is the one who would have acted anyway.

Propensity tells you who is likely to act. It can’t tell you whether your message caused it. Persuadability models estimate the difference contact makes, so budget goes to the people it can move and away from the people it can’t, or shouldn’t.

Would act anyway

Reaching them spends budget and claims credit for a decision they had already made.

Reached: acts · Not reached: acts

Persuadable

The only people for whom contact changes the outcome. This is where spend moves the number.

Reached: acts · Not reached: doesn’t

Leave alone

Contact makes them less likely to act. They belong on the suppression list, not the media plan.

Reached: doesn’t · Not reached: acts

Not this time

No message moves them in this window. Readiness may change; the score will see it when it does.

Reached: doesn’t · Not reached: doesn’t

Four kinds of people in any audience, sorted by what they do with and without contact. Illustration; proportions are not from a client build.

Calibration

A 0.7 should behave like seven in ten.

Gravity scores are calibrated probabilities, not rankings dressed up as percentages. Each published model is checked bin by bin on your holdout, and outcomes that return every day keep it honest after launch.

Illustration of a calibrated model. Validation reports show the same chart for your model on your holdout.

Model label

Every published model carries one.

Name and version: Named for the decision it serves, versioned and dated with the rebuild it came from.
What it estimates: One action, in one window, for one population. Nothing broader.
What it isn’t: Not an eligibility decision. Not a verdict about any one person.
Where it is weak: Thin-history records, new categories and small geographies are flagged, not hidden.
Evaluated against: A demographic baseline, a lookalike baseline and your current model, on your holdout.
Report: ROC and AUC, precision and recall, lift by decile, calibration, stability across rebuilds.
Permitted use: Marketing and engagement only.
If it lost: It was killed, and the reason is in the log.

Moonbrush-reported results from one full retrain cycle. Study details and the acceptance record are available under NDA.

What comes out

Not a score. A cohort your team can act on.

Five families of scores collapse into a card a planner, a strategist or a banker can read: who these people are, why they would act, when they can hear it and what to say. This one is a specimen from a financial services build.

Specimen cohort · Financial services · Consumer to business

The business that hasn’t told its bank yet

2.04M people paid for their own work, still banking as consumers. Some already bank with you.

How this cohort was assembled Specimen build. Counts refresh from the live graph.

270.2Mthe graph, rebuilt monthly
31.6Mobserved in category
7.8Mscored above threshold, five families
2.31Mseparated from three near neighbors
2.04Mactivation-ready, hashed

01 Built from

Behavioral eventsEvening invoicing and payroll research38
Intent topicsHow to get paid, never how to start11
Purchase cadenceSupplies at volumes no household uses5
Job profileA second title beside the employer’s4
Brand affinityInvoicing tools, not consumer fintech3

61 signals in five families. None is income, age, a balance or a product.

02 Why they act

MotivationBe treated as a business before the revenue proves it0.86
HopePersonal and business money finally separated0.77
FearBeing asked for paperwork they don’t have yet0.72
Won’t sayThat the first clients paid a personal account0.64

Decision style Weeks of private research after work. The account opens at the end of it.

03 When they can hear it

Near zero 09:00–17:00Peaks 20:30–23:30 weekdays

The category buys media hardest during business hours, while this cohort is still at work for an employer.

04 What to say

Message that landsOpen the account before the first invoice.

Message that misses: the category defaultBusiness checking with no monthly fee.

A fee was never the worry. The paperwork is, and so is being taken seriously before the revenue arrives. Message hypothesis, tested against your current offer in a controlled activation.

05 One household, eleven weeks

Week 1Side-income topic. First evening session.
Week 3Invoicing-tool research. Affinity starts to rise.
Week 4Supplies at retail, second order this month.
Week 6A second title beside the employer’s: consultant.
Week 8Payroll and tax topics. Score crosses threshold.
Week 10Business-banking research, 21:40 on a Tuesday.
Week 12An account opens. The ledger’s first sight of it.

The account lands with whoever spoke to this household in week ten.

Modeled demographics: Suppressed
Credit decisions: Never used
Match: Hashed, clean room
Measured on: Your holdout
Refresh: Monthly

Lines we don’t cross

What the models are never for.

The rules are written into the platform, not into a policy nobody reads. See how we handle trust.

Marketing and engagement only. Never used for credit, insurance, employment, housing or any eligibility decision.
Modeled ethnicity and political ideology are never selectable, never model inputs and never exported.
No data about minors, and no way to look up an individual.
Your data stays in its own instance, and identifiable client data never trains a model other clients use.
Identifiers leave as SHA-256 hashes. Raw personal information stays inside the platform.
Every model has to beat a demographic and a lookalike baseline on your holdout. Kill decisions are logged.
No security incidents or breaches to date, and a 72-hour notification commitment to clients.