"Don't just predict the future. Cause it."

WhyLab

Causal Decision Intelligence Engine powered by Multi-Agent Debate.
22-Cell pipeline bridging Causal Inference (Science) and Business Decision (Art).

Dashboard ↗ GitHub ↗

At a Glance

Pipeline Cells

AI Debate Agents

Benchmarks Validated

142

Tests

What it does

WhyLab answers "why?" — not just "what will happen?" It provides actionable causal verdicts: Rollout 100%, A/B Test 5%, or Reject. Three AI agents (Growth Hacker, Risk Manager, Product Owner) debate the evidence to reach a decision.

22-Cell Pipeline

Data — Discovery — AutoCausal — Causal — MetaLearner — Conformal
  Explain — Refutation — Sensitivity — Validation
    QuasiExperimental — TemporalCausal — Counterfactual
      Heterogeneity — Overlap — PolicyTree
        Viz — Debate — Export — Report
          Audit — Summary

Auto causal graph discovery (PC algorithm + LLM hybrid)
5 meta-learners (S/T/X/DR/R) with Oracle ensemble
IV, DiD, RDD quasi-experimental methods
Granger causality and CausalImpact time-series
Structural counterfactuals and Manski bounds

Benchmark Results

Validated on 3 standard causal inference benchmarks (10 replications each):

Dataset	Best Method	PEHE	ATE Bias
IHDP (n=747)	T-Learner	1.164 ± 0.024	0.039 ± 0.031
ACIC (n=4,802)	S-Learner	0.491 ± 0.017	0.018 ± 0.013
Jobs (n=722)	LinearDML	170.5 ± 32.3	39.2 ± 36.6

Multi-Agent Debate

Growth Hacker (10 evidence types) — finds revenue opportunities from causal signals
Risk Manager (8 attack vectors) — warns about potential losses and model vulnerabilities
Product Owner (Judge) — synthesizes Growth vs Risk — actionable verdict

Supports LLM-enhanced debate (Gemini API, MCP v2 protocol) with automatic rule-based fallback. 142 tests ensure reliability.