Seven signals, or one?

For a while the story looked like a chorus: half a dozen different signals on the watch all pointing the same way before a crash. Independent witnesses agreeing is powerful evidence. But when we checked whether they were actually independent, the chorus turned out to be mostly the same voice, recorded several ways.


The reduction

These are seven measured autonomic channels, some load-and-recovery, some orthostatic, and they don't measure seven different things. Grouped by how strongly they move together, they collapse into 3 roughly independent clusters:

The ties that collapse them

Each of these is a measured correlation. A value near +1.00 or −1.00 means two "separate" signals are really one (the minus sign just means one rises as the other falls):

What it does to the headline

Counting honestly (three or four independent clusters, not seven separate tests) only 1 clears conventional statistical significance at this sample size. Corrected honestly for roughly three independent clusters rather than seven separate tests, one cluster (the within-day stress spike) is the only one that clears the conventional one-in-twenty line in this multiplicity check. But be exact about what that does and doesn't mean. Judged the other way (on the whole four-year record at once, the single-pool test the scorecard actually uses) that same stress spike does NOT hold; it's a not-found on the scorecard. The signal that held up across the whole record is a different one: the autonomic-variability reading (sleep-stress stdev), which cleared the project's own pre-registered discrimination bar even though it sits below the stricter cluster-corrected line here. Two honest cuts of the same data, and they agree on the thing that matters: none of these is a predictor.

The findings are still real; "many converging channels" was the overstatement. So we changed the headline rather than the evidence, which is the whole point of checking.

The same fact, read the other way

Collapsing seven readings to one factor is a discount when you're counting tests. But for a different question (is the watch producing real signal at all?) the same fact points the other way. If the wrist were spilling noise, there'd be no reason for seven readings to fold down onto a single autonomic factor that moves coherently, in the right direction, and in proportion to how bad the day was. Noise doesn't organise itself that neatly. So the collapse doesn't make the signal weaker; it tells you what it is: one real autonomic signal, measured several ways, not seven, but not zero.

What this page can't settle on its own is whether that one factor tracks anything outside the watch. That rested on an external anchor, the felt-state score, and the one independently-known autonomic event in the record. That event has now been tested: during a 14-day COVID infection the factor moved significantly, a qualified pass, and tellingly, one carried by these same correlated stress and body-battery legs. So 'one factor, several ways' isn't only internal book-keeping; it's what moved for a real event. See the COVID check.

The correlation values shown are the real ones the research has reported; the full signal-by-signal matrix is still to be exported. Statuses and the exact effective-N may sharpen, but the collapse is the finding.

The full cross-channel correlation work, in the research repo: cross-channel-correlation.md ↗

Open full page ↗