Decision models, not language models.
The substrate runs randomized controlled experiments on synthetic respondents grounded in your customer data, then validates against published human studies replicated blind. Every reported ratio ships with its denominator and measured human ceiling.
Your customers, not proxies.
A panel built from your own customer data. No syndicated audiences.1
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.
A counterfactual, not a question.
We manipulate the decision and measure the outcome. Confounders are explicit before any run.2
Random assignment, in silico.
Treatment vs. control from your population. Each respondent sees one arm.3
Respondents are observed in the situation, not asked. Output is behavior, not a Likert summary.
Lift, ranked. Confidence, bounded.
Effect sizes with confidence intervals. Heterogeneous effects. Power against your field cohort.4
The falsifier.
If we lack priors for your cohort, predictions degrade. Every report carries an out-of-distribution flag.
The reviewed mean across 43 published studies is 0.73, with the human ceiling alongside. Read the audit log →
The methodologists.
Choice modeling (McFadden → Dubey). Experimental psychology (Moskowitz). Economics (Sekerke). The bench is installable via npx shadcn add (registry).
- De-identification follows the SOC 2 Type I controls. Re-identification probability ≤ 10⁻⁶ across the trait combinations we ship to the substrate.
- Subodh Dubey's choice-modeling framework, McFadden lineage. The DAG isn't decoration — it's the design matrix the simulation consumes.
- Language models on their own predict the next word. The substrate predicts the next action, conditioned on the trait vector and the situation prompt.
- 2,000 ongoing replications targeting the 2.5M-paper Subjective Probability database. Open audit log; the failures are public.