The rule everything obeys
This is the product, not an implementation detail.
data → deterministic analysis → evidence → conclusion → optional plain-English explanation
never data → language model → conclusion
A cause explorer that always produces seven explanations is a horoscope with citations.
Where a language model is involved at all — only in the Connected Coach, which is off until you turn it on — it may choose which analysis to run and rephrase what an engine concluded. It may never originate a conclusion. “Nothing meaningful found” is a result Brin is built to be able to report, and does.
Five ways to say how sure
Every claim carries one of these, and never anything stronger than the evidence supports.
| Vocabulary | What it means |
|---|---|
| Observed | Measured directly, on days the readings exist. |
| Similar timing | Two things moved close together. Not a cause, and the denominator is always shown. |
| Separate trend | Both are moving, but not together. |
| No detectable relationship | Checked and found nothing. A result, not a gap. |
| Insufficient data | Too few readings to check. Named, never silently dropped. |
The word “possible” is banned throughout. It implies a cause nothing has established.
In plain English
Five questions, asked of every measure. The methods below are how each one is answered.
- Is this a one-day shift, or a long drift?
- Is it outside your own recent pattern, or ordinary for you?
- Did anything else move around the same time?
- Have we seen the same response before?
- Is there enough data to say anything at all?
The methods
All deterministic, all tested, all on your phone.
| Engine | Method |
|---|---|
| Change-point | Pettitt’s test, with coverage guards, noise floors and a cap on how many changes may be claimed. |
| Change shape | Step versus slide, decided by residual comparison — so a four-month descent is never dated to one morning. |
| Trend | Mann–Kendall with tie correction for direction; Theil–Sen median slope for rate. |
| Baselines | Robust medians and MAD, per measure, from your own history. |
| Nervous system | Activation and restoration measured independently against your own distribution, split by context, nights divided into thirds. |
| Sleep | Stage union rather than the sum of overlapping provider records, plus need, debt and the overnight heart-rate dip. |
| Training load | Banister TRIMP with sex-specific weighting, plus whole-day cardiac cost. |
| VO₂ estimation | Eight field tests recognised inside workouts you already did — Cooper, 2.4 km, Rockport, Queen’s, Bruce, shuttle, 2000 m row, race. |
| Associations | Spearman with tie handling; sample counts carried through to the screen. |
| Interventions | Whether a thing you tried did anything, measured against your own days without it. |
| Experiments | Alternating-day or block self-experiments, with an end-of-period sweep across every measure. |
| Validation | Source inventory, duplicate linking, and a report you can read yourself. When one source misses a window another source’s readings can cover it — a window no source covered stays empty, and nothing is ever interpolated. |
A worked example
Activation and restoration are measured on separate signals — where your heart rate sits in your own distribution, and where your HRV sits in yours. Because neither is derived from the other, both can be high, both low, or one of each, and the total says which.
Built the other way, as two halves of one number, the pair summed to 100 on every real reading and the chart could only ever show a see-saw.

Searching twenty things at once
Checking twenty measures and reporting the one that moved is how coincidence gets published. Brin does four things about it, and showing the denominator is only the first.
The search space is visible
“5 of 20 moving” tells you how much was looked at to find it. This is transparency, not a statistical correction — it does not adjust any threshold.
A floor on what can be claimed
A measure needs at least a fortnight of readings before it is eligible at all, and the change-point test carries coverage guards and a noise floor beneath which nothing counts.
A cap on how many
At most three findings lead a morning. Past that a screen is a list of worries rather than a briefing, and a long list is itself a sign of noise.
Coincidence is never upgraded
Two things moving close together is reported as similar timing and labelled as such. It is never presented as a cause, and the caveat is printed next to it rather than in a footnote.
Similar timing could turn up by chance when checking 20 measures. This does not establish a relationship or a cause.