How we try to make the signal fail

The first question is never which row looks exciting. It is whether anything survives.

The formal side of this work proves properties of a definition. Whether the definition tracks a living body is a separate question, and an empirical one, and this note is about how we are trying to answer it — carefully, and against ourselves.

The test is a bridge: a longitudinal study that joins each DRTT scan to the same day's physiological signals from a wearable — the standard health rails, sleep timing, walking speed, sleep score, blood oxygen, and others — and asks whether the packet-native measurements our instrument produces move with any of them. It runs, at present, on a single subject over several weeks, roughly forty aligned scan-and-wearable days so far, of a target not yet reached. It is early. Nothing here is clinical validation, and nothing here is a confirmed result. What follows is the method, and an honest account of where it stands.

Why the first pass is deliberately strict

A single-subject daily bridge is one of the easiest places in science to find a signal that isn't there. The reasons stack up: the days are autocorrelated, so yesterday predicts today whether or not anything real is happening; the wearable signals are consumer-derived and sometimes smoothed; there are many packet metrics to test; the same physiological rails get reused across many comparisons; everything shares slow calendar trends; and the weeks may contain regime shifts. Any one of these can manufacture structure out of noise.

So the first question we ask is whether anything at all survives a screen built to destroy false structure. We made the first pass severe on purpose. A strict screen is what gives meaning to whatever passes it.

The survival gate

Before any relationship is allowed to count as a survivor, it has to clear five independent checks:

• Effective sample size. Because the days are autocorrelated, the true number of independent observations is smaller than the calendar count; a relationship has to hold on that reduced effective N.
• Circular-shift null. We slide the wearable signal in time relative to the scans and re-test. A real relationship weakens when its dates are pulled out of alignment; one that holds up under the shuffle was riding on the calendar all along.
• Trend robustness. We remove the shared slow drift from both series and require the relationship to keep its sign and a meaningful size afterward, so what remains is a real dependence with the calendar's common trend taken out.
• Differenced consistency. We require the day-to-day changes to move together as well, so what we are seeing is genuine movement in state, holding even once the standing levels are set aside.
• Sign and split-half stability. Two different correlation measures have to agree in sign, and the relationship has to keep its sign across the first and second halves of the data.

Only after all five does a relationship face a further, conservative correction for the sheer number of comparisons made.

The null behind the shift test is built to be honest about dependence. For each simulated replicate we shift each wearable rail once in time and reuse that same shift everywhere the rail appears — preserving the rail's own autocorrelation and the fact that it is reused across many comparisons, and breaking only the one thing under test: the alignment between scan dates and physiological dates. The question it puts is how many survivors would appear if the real physiological signals kept all their structure and were only decoupled from our scans — a far harder null than random noise, and the relevant one.

The preregistration boundary

There is one more discipline, and it is the one that separates honest tracking from wishful confirmation. The first stretch of days — the ones used to find the hypothesis — cannot also be used to confirm it. So the primary construct is preregistered with its exploratory window excluded, and confirmation requires a disjoint set of at least fifty aligned scan-days drawn only from later. Those discovery days did their job; they are not permitted to do it twice.

What survived, and what didn't

Here is the geography of the broad surface — the full set of relationships the screen was pointed at, grouped by scope:

Of those 5,487 relationships, 83 cleared all five rungs of the survival gate. After the correction for multiple comparisons, the number of broad individual relationships allowed to make a standalone claim was zero. A further 2,233 showed a tentative trend or a consistent sign and then failed a stricter rung or the correction — noticed, and not promoted.

Those 83 survivors are themselves within what the dependence-preserving null produces by chance: the global omnibus does not reach significance (p ≈ 0.17). The broad surface, in other words, is not globally coupled to the wearable. If the instrument and the wearable were simply sharing calendar drift or smoothing, that surface would light up everywhere. It does not, and that refusal is what makes the rest worth reporting.

Where signal does concentrate, it concentrates narrowly, and in physically coherent places. At the family level, the one grouping carrying a significant excess of survivors is pair-lag mismatch — the agreement of timing and lag between separated sites of the body. At the level of preregistered physical constructs, two survive the screen: site-timing / phase agreement — the organization of timing, lag, and amplitude across sites — tracks circadian and gait rails (mid-sleep timing, walking speed), across several related metrics and in day-to-day change rather than as one isolated correlation; and laminar sleep-recovery support — a route narrower than the broad laminar family, which is itself null and sits slightly below its null expectation — tracks sleep score, mid-sleep timing, and blood-oxygen minimum.

The shape of that is the point. A broad surface that mostly refuses to promote, with signal surviving only in two narrow, physically interpretable routes, is a far healthier result than a surface that correlates with everything.

What we are not claiming

Because the ways to over-read a study like this are well worn, the boundaries are worth stating plainly. This is not clinical validation. It does not show that DRTT diagnoses anything, or that the wearable proves the scan. It is not the confirmed preregistered finding — that requires the disjoint confirmatory window, which is not yet complete. The broad laminar family does not correlate with the wearable, and we do not claim it does. The honest description is narrower than any of those: an interim, preregistered tracking signal, surviving a strict artifact-discrimination screen at the construct level, with broad and per-link claims held under control, pending confirmation.

The road to confirmation

The next milestone is to keep collecting invariant-preserved daily scans until the disjoint confirmatory window reaches its fifty-day target — at the current pace, around early August 2026. At that point the preregistered confirmation runs exactly as specified, on data the hypothesis has never seen. Until then, this is what we have, stated at the strength the evidence supports: a strict first pass that mostly says no, and a narrow, coherent signal that has so far survived being told no.

The empirical companion to four notes on the formal side of the work: What is genuinely new, A boundary-observable certification algebra, The temperament of a bespoke algebra, and The making of an observable.