A long record lets you see changes that a week of tracking cannot. It also carries changing devices, missing signals and changing definitions. Start with the report’s static evidence, then use the toolkit project page to understand how an export becomes a report.

The case and its evidence

The report covers 13 February 2016 through 18 March 2026. Its public scatter contains 3,656 records across 2,963 distinct date values. Dates can carry multiple recordings, so the existing record-weighted means should not be described as independent nightly observations.

The overall quality mean is 83.3%; the 3,521 records without a travel tag average 83.7%. The 2020 annual mean is 94.4% (n=362). These are different denominators. A high consumer score does not establish healthy sleep.

Continuity of records is not continuity of measurement

Time in bed and estimated time asleep answer different questions. BD is a consumer breathing-interruption estimate, not clinical AHI. Morning mood is self-reported. The author’s adjusted scores add chosen weights; they are not validated clinical scales.

Metric coverage varies by year. The 2026 Jan–Mar BD mean is 6.88 with 78 valid BD records; March alone is 8.18, without a published monthly n. Several exercise and environment groups lack separate BD counts. Missing coverage must remain visible.

Sleep-stage discontinuities are a reason to investigate measurement changes, not proof of a particular app update. A comparison needs a stable definition, an explicit window and per-metric n before interpreting differences.

From an export to a report

The Sleep Toolkit project describes SleepCycle CSV input and structured JSON, interactive HTML and PDF output. The project page links to the web app and explains local use and session deletion. Those are published descriptions; this companion has not tested uploads, deletion timing or remote processing.

A useful first pass keeps the original export private, checks field names and units, and records the processing version with each result. The evidence page publishes an aggregate fingerprint for the existing chart payload. That fingerprint does not establish the historical processing commit or replace a raw-data reanalysis.

What the comparisons can support

Massage, caffeine, travel and exercise tags are overlapping observational groups. Their differences describe this person’s records and context. They do not isolate a treatment effect or generalize automatically to another person. Model counterfactuals remain predictions, rather than experiments.

The next methodological choice is how to assign sessions to a night and handle repeated recordings. A revised rule would change denominators and require recomputation. Until then, retain the published record-weighted results and their limitations together.