**The most counterintuitive finding is that late caffeine barely changes how long people sleep, but it clearly cuts their deep sleep. Exercise shows no link to sleep or fatigue at all.** All figures below are Pearson correlations across 8,500 users, from the correlation queries I ran on `raw.bedtime_screentime_sleep_debt`. - **Caffeine after 5pm vs. total sleep hours:** r = -0.04, essentially zero. A caffeine cutoff would not show up as more hours of sleep or a lower sleep-debt category. - **Caffeine vs. deep sleep %:** r = -0.34, the strongest caffeine effect in the data. Caffeine hurts sleep quality, not quantity. - **Caffeine vs. sleep latency and fatigue:** r = +0.20 for latency and +0.18 for next-day fatigue, so the effect on those is modest. - **Physical activity:** r = 0.004 with sleep hours and -0.03 with next-day fatigue. Despite the usual sleep-hygiene advice, activity minutes tell us nothing here. **Phone use dominates everything else.** Bedtime phone minutes correlate at r = 0.86 with sleep latency, 0.71 with fatigue and -0.55 with sleep hours. A relationship that tight is unusual for behavioural data. Fatigue also tracks sleep duration closely (r = -0.88) and alarm snoozes even more closely (r = 0.88). **Blue-light filters do very little.** Filter users (3,976 of the 8,500) had slightly shorter latency (39.6 vs 41.6 min) and slightly lower fatigue (3.58 vs 3.99). But the filter's correlations with latency (-0.06) and sleep hours (0.04) are tiny. This is despite filters being heavily marketed as a fix, while raw phone minutes matter far more. This data is observational, so none of these are proven causal effects. A caffeine cutoff would be justified by deep-sleep quality, not by sleep hours.