Write the hypothesis and agree on an exposure unit.
DECISION SCIENCE PLATFORM / EXPERIMENT LIFECYCLE
From experiments
to evidence.
Design better tests. Quantify uncertainty. Make defensible decisions.
ⓘ
Demonstration environment. All five preloaded experiments are synthetic. Statistical results are computed from the entered counts, but are not observations from real people or systems. This tool is not a causal-identification engine and does not replace research design or scientific review.
EVIDENCE PORTFOLIO
Experiment outcomes
CROSS-PROJECT VIEW
Estimated treatment lifts
DECISION REVIEW
Portfolio signals
Choose a detectable effect, alpha and statistical power.
Review allocation balance and expected event counts.
Interpret effect size, confidence interval and limitations.
ACTIVE REGISTER
Experiment portfolio
| EXPERIMENT | SECTOR | EXPOSURES | RELATIVE LIFT | TWO-SIDED P | INDICATION |
|---|
A statistically detected difference does not establish causality without valid assignment, measurement, and protocol integrity.
Statistical analysis
Compare two independent conversion proportions. Calculations update from stored experiment counts, not fabricated model predictions.
RATE COMPARISON
Observed conversion rates
EFFECT ESTIMATE
Absolute treatment difference
METHOD GUARDRAILS
Diagnostic checks
BUSINESS TRANSLATION
Illustrative unit economics
Sample-size planner
Approximate the sample size needed per arm for a two-sided independent-proportions z test, assuming equal allocation, fixed design and the specified uplift.
BEFORE YOU LAUNCH
Design parameters
RECOMMENDED DESIGN / APPROXIMATE
Do not repeatedly peek at interim p-values and stop when significant without a pre-specified sequential-testing procedure. Formula assumes independent observations.
Scenario simulator
Explore the sampling variability implied by the observed rates. The simulation uses a seeded normal approximation, not real-world future events or a Bayesian posterior.
SAMPLING VARIABILITY
Simulated treatment effects
These simulated differences reflect estimated variance under a simplified model. The proportion above zero is not the probability that the treatment truly works.
Reports & data
Export your projects for reproducibility and transparent review. Saved work stays in this browser unless explicitly exported.
EXPORTS
Download evidence
Keep analysis results, experiment definitions, warnings and the export timestamp. Report conclusions are explicitly conditional.
IMPORT / RESET
Restore a workspace
JSON imports replace current experiments after confirmation. Existing work is never deleted automatically.
EXPLICIT LIMITATIONS
Interpret with care
What this provides
Independent two-proportion z test, Newcombe-Wilson confidence intervals, allocation checks, power approximation, seeded resampling approximation, and exportable documentation.
What it does not establish
Randomization quality, absence of contamination, treatment compliance, unbiased outcomes, external validity, multiple-test adjustment, or a causal effect by itself.
Quality before significance
Pre-register a primary metric and stopping rule, validate instrumentation, examine guardrails, and investigate sample-ratio mismatch before deciding.