The substrate runs randomized controlled experiments on synthetic respondents grounded in your customer data, then validates against real human behavior. 350+ replicated studies, 93% bioequivalence, sub-five-minute turnaround.
A panel built from your own customer data — de-identified histories, product traits, journey events. We don't run on syndicated audiences. Your model trains on your population.1 What that means in practice: the synthetic respondents are statistically indistinguishable from your real customers on the dimensions that drove their last decision.
Connection takes between four and eight hours. The substrate retains conditional probability structure across attributes — income, channel preference, prior conversions, complaint history — so simulated cohorts behave like real ones.
The decision is the variable we manipulate. The outcome is the metric we measure. Confounders are the things that would make us draw the wrong conclusion if we squinted at correlations. The methodology forces all three explicit before the simulation runs.2
Treatment vs. control, drawn from your population at the cohort sizes that match your downstream stat-test. Each respondent sees one arm. Every action — a click, a skip, a price-sensitive flinch — gets logged with the standard battery of choice indicators. This is the part that fails on language-model-only stacks.3
The synthetic respondent isn't asked what they'd do. They're put in the situation and observed. The output is behavior with metadata, not a Likert summary. The doctrine — stated vs. revealed — is the methodology, not the marketing.
Causal effect sizes with confidence intervals. Heterogeneous treatment effects where the cohort splits the answer. Power calculations against the cohort sizes you'd need in the field. We tell you what would happen, what wouldn't, and where the residual uncertainty hides.4
The substrate fails when the population trait distributions in the connected data don't span the decision space you're asking about. If you ship to a cohort we don't have priors for, our predictions degrade. We surface this case explicitly: every report carries an out-of-distribution flag, and we'd rather you collect 200 real respondents than ship a number we can't trust.
We'd see the bioequivalence vs. published-study replications drop below 80% on a methodology audit. That's the line we hold. The current rolling number is 93%; it has not crossed 80% in eighteen months. Read the audit log →
A choice-modeling lineage (McFadden → Dubey), an experimental-psychology lineage (Moskowitz, the "spaghetti sauce" Harvard line), and an economics lineage (Sekerke, Durham PhD / UChicago MBA). The substrate exists because these people read the same papers and disagree about them in the same room. The bench, mounted live below, is the same roster control consuming apps can install via npx shadcn add.