Interactive field lab · Data provenance
Nutrinaut Data Forensics Lab
Before trusting a trend, inspect what each row actually represents. Explore calendar shifts, workout duration definitions and possible duplicate imports with a transparent, browser-local check.
A number needs a source, a time and a definition
A 45-minute workout, a 60-minute clock interval and a 1,080-minute daily coverage window may all be valid fields. They do not measure the same thing. In a technical audit of Nutrinaut’s training-data paths, we found that session records and daily rollups have different meanings. We also found that a processing timestamp cannot substitute for the original event date. Those observations motivated the checks below; the downloadable sample is invented, not a user export.
- Preserve the recordKeep the source, full timestamps with UTC offsets and the kind of row.
- Compare definitionsSee elapsed, wall-clock, reported and moving minutes separately.
- Keep a receiptExport the rows, flags and limits before changing a diary or sharing a support report.
This is a data-quality aid, not a diagnosis, calorie estimator or device-validation test. A possible duplicate is a question to investigate, never an automatic deletion.
Try the calendar and duration check
Start with the synthetic sample or select a small, redacted CSV without names or sensitive notes. This checker reads the file locally and does not upload CSV rows to Nutrinaut or a model provider. It stores nothing in your account.
Columns: id,kind,source,start,end,reported_minutes,moving_minutes,processed_at. Kind is session, sleep_interval or daily_window. Timestamps need an explicit offset, such as 2026-10-11T10:00:00+02:00. A blank minute value means missing; 0 means a recorded zero. Limit: 250 rows and 128 KiB.
What the demonstration reveals
One ride, two imports
The watch and health bridge each show a session from 10:00 to 11:00. The lab marks a possible duplicate pair; it does not remove either. Its 45 reported and 42 moving minutes may use different pause rules from the 60 elapsed minutes.
A window is not a workout
The daily_window row runs from midnight to 18:00: 1,080 elapsed minutes. It represents daily coverage in this demonstration. Adding it to a 45-minute session would make an invalid workout total.
The repeated autumn hour
From 23:00 UTC+02 to 07:00 UTC+01, the local clock labels span 480 minutes while the actual interval spans 540 minutes. The offset explains the difference; it does not establish minutes actually asleep.
What about exactly 15 minutes? A source may contain a measured, manually entered or assumed 15-minute duration. Without a provenance field, this lab will not guess which. In the inspected Nutrinaut training path, an unspecified duration can be represented by a 15-minute assumption, but a stored value of 15 alone does not prove that happened for that record.
For an overnight interval, compare exact timestamps before moving its diary date. See One Night, Two Dates. For workout energy, check the underlying minutes before using an estimate: Before You Trust a Workout Calorie Estimate.
Metric passport · Synthetic sampling case
A change in sampling can change a summary
Before combining an HRV series across a watch or software change, record the exact metric name, source, recording conditions and change date. Apple now describes separate overall and Recovery HRV variants. Its documentation says Recovery HRV is sampled while you are still, so measurement counts vary with activity. These manufacturer descriptions do not establish that two exported fields are interchangeable.
Invented example, not a person’s measurements. Suppose one comparable recording window averages 40 ms and another averages 80 ms. Giving the second window more readings changes a pooled reading-weighted mean, even though neither window’s mean changed.
Neither weighting rule is automatically the correct personal baseline. Keep unlike variants separate, preserve recording context, and mark a device or software change before interpreting a trend. This demonstration does not assess recovery, diagnose a condition or recommend whether to train.
Use the metric passport in the Apple/Garmin comparison. Sources: Apple’s 9 September 2026 announcement; Apple’s recording-condition explanation. Method: Sampling demonstration 1.0 · Eugene Blago · 15 October 2026.
Method, limits and sources
Calendar/duration method 1.0; shared-formula demonstration 1.0; sampling demonstration 1.0. This page parses explicit-offset ISO timestamps; elapsed minutes come from the two instants, while wall-clock minutes use their printed local clock fields. A candidate duplicate means two session rows start and end within two minutes of one another. Those are disclosed interface rules, not validated error rates. No record is merged, corrected or transmitted. A different offset can reflect travel as well as a seasonal clock change; without a time-zone identifier and source history, the page cannot distinguish them.
Checks cover only the columns shown above. They cannot verify a wearable sensor, infer missing sleep, decide which of two records is correct, or establish a person’s training load. Historical Nutrinaut records informed the failure-mode selection, but the public sample and all figures on this page are synthetic. We have not measured a population error rate or validated this tool against clinical outcomes.