Most self-tracking answers are wrong. This one argues with itself.
It takes one person's own exports — a glucose sensor, a
sleep ring, a chat log — answers the question they actually asked, and then
spends most of its effort trying to prove that answer wrong before showing
it to them.
1studies published
47claims under test
Libre 3 · Oura · chatdata sources
How it works
Running the tests is the easy half. Refusing to believe them is the
product.
01
Register the question
The question is written down, with its outcome and its exposure, before the data is opened. That is what separates one confirmatory test from the hundreds of exploratory ones a dataset will happily offer you.
02
Audit before analysing
Empty fields in live templates, dead checkboxes, coverage gaps — and informative missingness, because people log least on their worst days, and every complete-case test quietly inherits that tilt.
03
Test, then attack
Surviving a significance threshold earns a finding nothing but a turn on the stand. Seven passes then try to kill it: drop a month, reverse the arrow, adjust for the obvious third cause, ask whether the exposure is just what a bad day looks like.
04
Report the verdict, including nulls
HOLDS, FRAGILE, CONFOUNDED, ARTIFACT or NULL — assigned by rule rather than by taste, and reported the same way whether or not the answer is the interesting one.
Published studies
One person, one question, every number recomputed from the raw export each
time the page is built — so the prose cannot drift from the result it
describes.
Bob — one of our early clients · 2024-01-01 → 2024-11-01
“look at dawn glucose level if consistently higher than 100 mg/dL”
Strongly clock-aligned. The average day runs 26.4 mg/dL from a 04:00 trough to a 13:00 peak, and the dawn rise averages +19.4 mg/dL. His own 2024 question — is dawn glucose consistently over 100 — comes back “usually, and by about five”: 104.9 on average, clearing 100 on 67% of days. It survives dropping any wear block and dropping every sensor's first 24 hours.