**Overall, the data is clean, with one real inconsistency: the sleep-debt categories overlap on hours.** I found no nulls, duplicates, impossible values, mixed units or format variants. **What checks out** - **Completeness and keys:** All 8,500 rows have a distinct `user_id`, and every profiled column has 8,500 non-null values (step-0 profile). - **Ranges are plausible and show no impossible values (step-0):** - Age runs 18–65. - Screen brightness is 10–100%. - `blue_light_filter_active` is strictly 0/1. - Caffeine after 5pm is 0–250 mg. - Physical activity is 0–112 min. - Alarm snoozes are 0–7. - Fatigue is 1–10. - Total sleep is 3.2–9.8 hours. - Phone minutes at bedtime are 1–180. - **Units look consistent:** Deep-sleep % (8.1–28) and REM % (9.6–27) are on the same percent scale. No row has deep + REM above 100 (`deep_rem_over100` = 0). - **Categorical labels:** Counts for gender, chronotype, occupation, app and blue-light flag (step-1) show no spelling or case variants. The `user_id` check found no whitespace problems (`ws_ids` = 0). **Issues and oddities** 1. **`sleep_debt_category` is not a clean function of `total_sleep_hours`.** - Mild Deficit (6.75–7.74 h) and Optimal Recovery (7.75–9.8 h) split cleanly on hours. - Severe Sleep Debt (3.2–5.38 h) and Moderate Debt (3.2–6.74 h) overlap heavily (step-3). - Of the 4,462 Moderate rows, 207 sleep under 4 hours. 440 of the 647 Severe rows are also under 4 hours. - Severe rows also have much longer sleep latency: average 72.7 min versus 42.4 for Moderate, and 104 of them exceed 90 min versus 13 for Moderate (step-2). - The category was meant to be derived from hours alone. Either it also uses latency or another input, or labels are inconsistent. Don't re-derive or rely on it as an hours-based band until the rule is confirmed. 2. **Possible floor and ceiling clipping.** Total sleep hits exactly 3.2 as the minimum in both Severe and Moderate, which suggests a floor. Caffeine tops out at exactly 250 mg, and bedtime phone minutes run exactly 1–180. I didn't test whether values pile up at those limits, so treat this as a suspicion. 3. **Zero-inflated caffeine.** Summing the `zero_caf` counts in step-2 gives 4,950 of 8,500 rows (about 58%) with 0 mg after 5pm. That could be genuine non-consumers or a default fill for missing data, so be careful when modelling caffeine effects. 4. **Long-tail sleep latency.** The maximum is 123.3 min and the minimum is 6.0 min. The 99th percentile is 114 min for Severe and 82 min for Moderate (step-2). These are plausible but extreme, and they are concentrated in the Severe group. 5. **Small and imbalanced groups.** Non-Binary has 248 rows, and News/Reading has 596 against 2,198 for TikTok/Reels. Estimates for these groups will be noisy. This is a representativeness issue, not an error. 6. **Sensitive field.** `bedtime_phone_minutes` is flagged as person-identifying, so report it only in aggregate. **Recommended next step:** Confirm how `sleep_debt_category` is generated, and check whether values pile up at 3.2 hours, 250 mg and 180 minutes. That would show whether those limits are caps or genuine measurements.