I pulled three samples from `raw.bedtime_screentime_sleep_debt` (8,500 rows, one per `user_id`), and each is shown as a rendered table above. I left out `bedtime_phone_minutes` because the card flags it as person-identifying, and I left out the ingestion metadata columns. **What the samples cover** - **Random draw:** two draws of 15 rows each (the first and the third) show the typical mix of occupations (Corporate 9-to-5, Healthcare / Shift Worker, Remote Tech, Student, Freelance / Creative), chronotypes (Morning Lark, Intermediate, Night Owl), genders (Female, Male, Non-Binary) and all six bedtime apps. - **Stratified draw:** the second sample takes 4 rows from each `sleep_debt_category`, so the rare Severe Sleep Debt group is visible, not just the common Moderate Debt group. **How the full dataset compares** - Moderate Debt is 52.5% of rows (4,462), Mild Deficit 23.6% (2,004), Optimal Recovery 16.3% (1,387) and Severe Sleep Debt 7.6% (647). A purely random sample will therefore be mostly Moderate Debt, which is why I added the stratified one. - The category follows sleep duration. The average is 8.35 hours for Optimal Recovery, 7.23 for Mild Deficit, 5.55 for Moderate Debt and 3.76 for Severe Sleep Debt. - Snoozes and fatigue rise as sleep falls. Average snoozes go from 0.81 (Optimal Recovery) to 5.88 (Severe), and average fatigue from 1.08 to 9.58. - In the sample rows, Severe Sleep Debt cases show long sleep latency (61–84 minutes), as in USR-02124 and USR-06860. Optimal Recovery cases show fatigue of exactly 1.0. - Some columns have heavy zeros. Many rows have `caffeine_post_5pm_mg` of 0 (the median is 0 in three of the four categories), and `blue_light_filter_active` is a 0/1 flag. I can also pull a sample filtered to a segment, such as Night Owls or one app, if that would be more useful.