summary_html = f"""

Question: What patterns exist in the data?

Answer: Training and complaints have correlation r = {corr:+.2f}

Conclusion: ❌ "Training increases complaints" (WRONG!)

Problem: Confounding by the hiring wave creates a spurious correlation

Question: If we INCREASE training, what happens to complaints?

Method: Backdoor adjustment (control for tenure)

Answer: Each training point REDUCES complaints by {abs(effect_adjusted):.2f}

Conclusion: ✅ Training actually HELPS (the dashboard was backwards!)

Question: On {worst_day['date'].strftime('%B %d')}, if training had been +10 points, what would complaints have been?

Method: Abduction → Action → Prediction

Answer: Complaints would have been ~{abs(counterfactual_analysis(worst_day_idx, 10)['change']):.0f} fewer

Power: Individual-level attribution for specific incidents

The CEO's dashboard showed: Training ↑ → Complaints ↑ (cut training!)

The causal analysis showed: Training ↓ → Complaints ↑ (increase training!)

The counterfactual showed: +10 training on worst days would have avoided {total_avoided}+ complaints

🚨 Dashboards show correlation. Decisions require causation.

""" display(HTML(summary_html))