The Consumer World Model.
Two models. One complete picture of your members.
Your AI shouldn't decide from a spreadsheet. It should decide from a model of your base.
A Behavior Model learns how your members actually behave — trained on their own event stream, per brand, never pooled. A Digital Twinsimulates what they'd do next — playing campaigns forward before you send them. Together they form a complete, living model of your member base: the thing every decision in the Decide node is made against.
Senses
AI-native core
Agent layer
Model layer · Consumer World Model
Behavior Model
learns what members do
Digital Twin
simulates what they'd do
Acts
↻ every outcome flows back in — the model keeps calibrating on your own data
Two halves of one model
What they did. What they'd do.
A world model of your consumers needs both directions of time: a faithful memory of real behavior, and a credible rehearsal of behavior that hasn't happened yet. One without the other is half a picture.
Behavior Model
A model of how your members actually behave.
A compact behavioral model trained on your members' own event stream — visits, purchases, points, opens, clicks, dozens of governed event types — one model per brand, never pooled across customers. It learns each member's behavioral fingerprint and keeps a live read on what tends to come next: purchase or lapse, and when.
- Member embeddings — a behavioral fingerprint per member — the engine behind lookalike audiences
- Next-event prediction — purchase and churn probabilities that feed segments and the agents
- Timing — each member's next-purchase window and the hour they actually engage
Honesty gate:the model only goes live for your brand after it beats a transparent baseline on your own held-out data. Until it earns that, the platform runs on glass-box math — it never fakes a model it doesn't have.
Deep dive: Behavior Model →Digital Twin
A simulation of what your members would do next.
The same model, run forward. The digital twin plays a plan against your member base before reality does: a campaign versus deliberately doing nothing, a journey change versus the status quo — member by member, rolled up into an honest forecast, calibrated against held-out reality.
- Counterfactual rollout — simulate the campaign and the silence, and read the difference
- Held-out calibration — forecasts are checked against reality the model never saw
- Research twins — de-identified member twins you can pre-test surveys and concepts onEarly access
Honest by design:a simulation is a rehearsal, not a promise. The twin reports confidence with every number and stays silent where your data can't support an answer — it will tell you “not enough signal” before it tells you a story.
Deep dive: Digital Twin →Benchmarked in the open — against the classical playbook
Public datasets · strict time-cut holdout · every method graded on the same unseen future
+6 AUC points
Next-buyer ranking vs hand-crafted-feature ML
On a public e-commerce behavior benchmark of ~100M real events, the behavior model ranks tomorrow's buyers ahead of classical machine learning by six AUC points — and ahead of every frequency heuristic it was raced against.
+2.5 pts top-1
Next-category prediction, statistically backed
On a public two-year retail-loyalty benchmark, next-category accuracy (a 128-way call) beats the classical-ML rung with the full confidence interval clear of the bar.
The only line that rises
Grew the data 5× — only the model improved
Classical methods have structural ceilings: counting doesn't get smarter with more data. When we scaled the benchmark corpus five-fold, the behavior model was the only method whose score went up.
Benchmark results on public datasets, under the same discipline the platform applies to your brand: models train only on the past, are scored on a future they never saw, and nothing serves until it beats the baseline on your held-out data. Public numbers prove the method — your gate proves it on your data.
One world model, two halves
The ontology models your business. This models your members.
Flash's Business Ontology gives AI a world model of the business — what a customer is, what has happened, what an agent may do about it. The Consumer World Model completes the picture with the part no schema can hold: how your members behave, and what they'd do next.
Together they're why a Flash agent's call is more than a lookup: it knows the facts through the governed semantic layer, and it knows the people through the model — before it decides anything.
Business Ontology
The nouns & verbs
objects, relationships, governed actions
Consumer World Model
The behavior & futures
what members do · what they'd do next
Every AI decision in the Decide node
SoClaw · Intent → Campaign · suggestions · insight
Where it shows up
Four decisions the model changes.
You never operate the model directly. You feel it in the quality of the calls the platform makes — the audiences it finds, the members it saves, the campaigns it rehearses, the moments it picks.
Find more members like your best ones.
Pick a handful of proven members — top spenders, a winning segment, even a single VIP — and the behavior model finds the members whose behavioral fingerprint sits closest to theirs. Not “same age, same city”: same rhythm of visits, categories and responses, read from what they actually do.
See churn forming — and act while it's still cheap.
For each member the model keeps a live read: how likely their next purchase is, how likely they are to lapse, and what's at stake if they do. When the risk crosses your line, the decision goes to the agents — SoClaw picks one fitting gesture (or deliberately none) inside your guardrails, proven against an untouched control.
Play the campaign forward before you send it.
The digital twin runs your campaign against the model of your base — and runs doing nothing as the counterfactual — so you see the expected difference before a single message goes out. Forecasts are calibrated against held-out reality, and the model says so when it doesn't have enough data for an honest answer.
The right member, at their own right moment.
Each member has a purchase rhythm and an hour of day when they actually engage. The model learns both — when the next purchase is due, when this member opens and clicks — so sends land in each member's window instead of everyone's Tuesday 10am.
Capabilities, benchmarked
Every capability earns its place. The winner serves it.
Behind the model is a capability benchmark: each prediction is tested against the strongest traditional method on held-out data, and whichever wins, serves. Where the sequence model wins, you get it. Where classical statistics win — timing rhythms, spend forecasts — those serve instead, honestly labelled. No capability ships on faith.
Consumer state
Sequence modellifecycle, intent, engagement — the live read every decision starts from
Churn likelihood
Sequence modelwho is about to lapse, 30-day horizon — feeds win-back automatically
Purchase likelihood
Sequence modelwho is about to buy — short-list refined by forward simulation
State change alerts
Sequence modelthe moment a member breaks their own pattern — the strongest early signal
Purchase timing
Rhythm statisticsexpected next order date and the hour each member actually engages
Customer value
Classical fitexpected 12-month value in dollars — served by the proven classical fit
Open to your own AI · MCP
Your commerce stack's AI can call it too.
The Consumer World Model isn't locked inside Flash. Its predictions are exposed as MCP tools— so the AI assistants your team already uses, or your own e-commerce agents, can ask about a member before they act: recommendation engines checking purchase likelihood, service bots checking churn risk before offering a save, ops copilots pulling a member's state mid-conversation.
- One scoped key — a read-only predictions scope — the AI sees predictions, never raw PII
- A governed envelope — every answer carries its confidence basis, freshness, reasons and limits — the AI can explain a number but can't overclaim it
- Honest absence — a member the model can't score returns nothing — never a guessed zero
Any MCP client · Claude, Cursor, your own agent
> "Is member m_8f2 worth a retention offer?"
→ cwm_predict_churn 82/100 · fresh tonight
→ cwm_predict_customer_value $412 / 12mo
→ cwm_predict_timing next order ~Aug 20 · sends land at 7pm
"High churn risk on a $400+ member — worth a
save. Their order window opens in two weeks;
schedule the offer for the evening."Tools: cwm_get_consumer_state · cwm_predict_churn · cwm_predict_purchase · cwm_predict_timing · cwm_predict_customer_value — one line of MCP config, scoped API key, per-call audit.
Recursive self-improvement · governed
A model that grades its own predictions — and retrains on the verdicts.
Every prediction the model writes is snapshotted the moment it's made. When its horizon passes, reality grades it: did the member actually buy, actually lapse? Those graded outcomes become the training labels for your brand's own model delta — trained on top of a shared base that stays frozen. The improved delta serves only after it beats the base on your own held-out data. Predict, observe, grade, retrain, earn the right to serve — a loop that compounds, under governance.
Predict
scores written to members & segments
Observe
real orders and silences arrive
Grade
matured predictions scored against reality
Retrain
verdicts become your delta's labels
Earn to serve
beats the base on holdout — or doesn't ship
…and the improved model's next predictions enter the same loop
Your loop, your model only
The loop trains your tenant's delta on your outcomes. The shared base is frozen — no other brand's loop touches your model, and yours touches no one else's.
Causally honest grading
A prediction that triggered an action — a save offer sent because churn looked high — is excluded from its own accuracy score. The loop learns from clean evidence, not self-fulfilling prophecies.
Worst case: status quo
A retrained delta that can't beat the shared base on your held-out data never serves. The loop can make your model better; it is structurally unable to make it worse.
Why you can trust it
A model you can hand decisions to.
A world model of your customers is powerful — which is exactly why it runs inside the strictest rails on the platform. These aren't policies; they're enforced by the machine.
Yours alone
One model per brand, trained only on that brand's members. Nothing is pooled across customers; your model never learns from anyone else's base — or teaches theirs.
Earns its job
The model activates only after beating a transparent baseline on your own held-out data — and keeps being re-verified. If it stops earning its place, the platform falls back to glass-box math.
Forgets on request
When a member is erased, their data leaves the model too — retraining and purging are part of the deletion machinery, not a manual afterthought.
Honest at the edges
Predictions carry confidence; simulations are calibrated against held-out reality; research twins are built from de-identified profiles. Where the data can't support an answer, you get silence, not fiction.
The predictions the model feeds are part of the same glass-box discipline — Predictive Intelligence →
How agents act on this state — every skill declares the CWM state it needs before it runs: Enterprise Skill Library →
The full argument — including a ten-question buyer's checklist — is in the whitepaper: The Consumer World Model (PDF) →
The model surfaces where you work: member profiles, segments, campaign pre-flight and the agents' reasoning.