I pulled two samples, shown in the data blocks above. Both leave out `bedtime_phone_minutes` (flagged as person-identifying) and the ingestion metadata columns. **1. Random sample (15 rows).** This is a plain draw across the table. - Rows are mostly Moderate Debt or Mild Deficit, which fits the overall mix: Moderate Debt is 52.5% of the 8,500 users, Mild Deficit 23.6%, Optimal Recovery 16.3% and Severe Sleep Debt 7.6%. - Corporate 9-to-5 dominates this draw, and Healthcare / Shift Worker, Freelance / Creative and Remote Tech also appear. - Chronotypes (Night Owl, Intermediate, Morning Lark) and bedtime apps (TikTok / Reels, YouTube, Instagram / Reddit, Messaging / Chat, Streaming, News / Reading) are all represented. **2. Stratified sample (16 rows, 4 per `sleep_debt_category`).** A random draw rarely shows the rare Severe category, so this one makes the categories easy to compare. The category averages come from the table-wide aggregate (step-2): | Category | Avg sleep hours | Avg alarm snoozes | Avg fatigue score | |---|---|---|---| | Optimal Recovery | 8.35 | 0.81 | 1.08 | | Mild Deficit | 7.23 | 1.81 | 1.83 | | Moderate Debt | 5.55 | 3.35 | 4.68 | | Severe Sleep Debt | 3.76 | 5.88 | 9.58 | - The sampled rows follow the same pattern. Optimal Recovery rows have about 7.8–8.6 hours of sleep, 0–1 snoozes and a fatigue score of 1.0. Severe rows have 3.2–4.0 hours, 4–7 snoozes and fatigue of 8.8–10. - Sleep latency is also higher in the Severe sample rows (61–84 minutes), against 28–44 minutes in the Optimal Recovery rows. - Ranges seen in the samples: age 18–65, screen brightness 10–86%, caffeine after 5pm from 0 to 177 mg, and physical activity from 0 to 79 minutes. `blue_light_filter_active` is a 0/1 flag. `sleep_debt_category` is derived from `total_sleep_hours`, so the tight link between sleep hours and category is by construction. The links to snoozes and fatigue are real in this data, but this is observational data, so they don't show cause.