John S set Oura → Google Band → Sleep Study → 8 Sleep. John then asked for the normalised sleep study FIRST, on the grounds it is a clinical device carrying strictly more information (AHI, SpO2, fragmentation). Built John’s order. Note the premise did not survive testing: there is no common scale to put them on (see below), so this choice decides which device reports a night AND in whose units — it is not just a preference between equivalents. Worth one sentence between them.

Fitted the candidate curves against the 45 paired Oura/sleep-study nights before trusting any of them. Correlation between the two devices is r = 0.195; every one-parameter model (ratio 1.13, ratio 1.10, offset +0.68h) has NEGATIVE R², i.e. predicts worse than simply guessing the mean, and the best-fit line has slope 0.16 — what the study measured barely tells you what Oura would have said that night. They agree within 30 min on 16 of 45 nights. So the correction (0.68h) is smaller than its own error bar (0.88h) and adds error. Nights are now shown as the device reported them, with the source named. What IS stable is the LEVEL — Oura averages +0.68h over the study — so an aggregate mixing sources can be level-corrected even though a single night cannot be translated.

A source reporting under 3h while another saw a full night happens 28 times for Oura, 5 for the sleep study, 1 for Eight Sleep. On 2026-08-24 Oura logged 18 minutes against Eight Sleep’s 6.0h and that 18 minutes was the number on this page. Those are failed recordings, not short nights, and no scale factor rescues them. The reference scale being the least reliable capturer is fine — Oura defines the units, it does not have to supply the value.

First worn 2026-08-29. No rows in health.db yet, and no fetcher exists — the other devices each have one (fetch_oura.py, fetch_eightsleep.py). Two routes: the Fitbit Web API, which matches the existing pattern but needs an app registration John must do; or letting the Fitbit app write into Apple Health, which needs no new code but arrives unlabelled and so cannot be calibrated separately. It needs ~20 paired nights against Oura before its readings can be trusted on the same scale.

Set and rep detail is the one thing the adherence integration cannot get today. Nothing to build until the key lands; then strength can show prescribed-vs-completed per set.

John S can import them into the Bridge program from an upload, screenshot or rough recall, so the blocker is getting sessions recorded at all. The last one that reached the database was the manual “Gym with Ben” entry on 2026-08-11.