**The most surprising finding is how clean and strong the relationships are, not any one reversal.** Daily social-media hours track nearly every wellbeing measure, which is rare in real observational data on people. - **Usage vs. mental health:** the correlation is **-0.85** across all 4,500 students. That means daily usage alone explains roughly 72% of the variance in the Mental Health Index (r² ≈ 0.72). - **Other outcomes:** usage correlates at -0.72 with sleep hours, -0.71 with GPA and -0.68 with sleep quality. It correlates at +0.75 with perceived stress. - **Downstream links are also strong:** sleep vs. mental health is +0.77, stress vs. mental health is -0.77, and GPA vs. mental health is +0.66. - **The Overall_Impact groups form a steady gradient.** Average daily usage rises from 4.4 hours for "Beneficial" students (3,681) to 8.5 for "Neutral" (654) and 12.25 for "Negative" (165). Over the same groups: - Mental health falls from 84.3 to 63.4 to 47.1. - GPA falls from 3.55 to 3.00 to 2.54. - Sleep falls from 7.03 to 5.48 to 4.23 hours. - Stress rises from 11.2 to 22.3 to 31.1. - There are no groups that break the pattern. **Where correlation vanishes.** Weekend extra hours are essentially uncorrelated with daily usage (r = -0.01), and age is essentially uncorrelated with GPA (r = 0.03). Heavy weekday users are not heavier weekend users, which is itself odd. **Interpretation.** The dataset's own framing is organic and observational, but correlations this tight (-0.85 from a single behavioural variable) usually point to simulated or rule-generated data, or to a heavily engineered scoring scheme. Treat the results as a pattern in this dataset rather than evidence of real-world effect size. They also cannot show causation, since heavier use could follow from poor mental health as easily as cause it. I did not test this against the individual segments (platform, gender, late-night use). I'd check those next to see whether the gradient holds within each one. If it holds almost perfectly everywhere, that would strengthen the synthetic-data reading.