A discriminator, not a predictor.
This is the single most important idea on the site, and the easiest to get wrong. A signal can clearly tell crash-weeks apart from ordinary weeks after the fact, and still be almost useless as a warning. Both things are true at once. Here's why.
Two different questions
Discrimination asks: looking back, did this signal run differently in the days before a crash than on an ordinary day? Several of ours do; that's a real result. Prediction asks something much harder: if the signal fires tomorrow, will a crash actually follow? The gap between the two is created by one stubborn fact: crashes are rare.
What happens when it fires
Take a generous version of our best signal across 1,000 days. It fires on about 687 of them. But because crashes are so rare to begin with, only about 18 of those fires are actually followed by a crash. The other 669 are false alarms.
| days it fires | a crash follows | false alarms | |
|---|---|---|---|
| per 1,000 days | 687 | 18 | 669 |
So even at its best, a fire is followed by a crash about one time in 37, roughly a 2.7% hit rate. The single signal that held up across the whole record (HA07d), at its real numbers. When it fires, a crash follows ~1 time in 37, wrong the other ~36. Every other tested signal sits at a lower or equal lift. This reads the body; it does not predict it. A weather report, not an alarm.
This shift, from predicting a crash to understanding one, is where the whole investigation was heading. Why we started simple, and only later reached for complexity →
The honest ceiling, as a ladder
Findings are sorted not by how exciting they look but by how far they actually reach:
- Tier A Predictive No signal here reaches this.
Forward-validated: when it fires, a crash reliably follows. Could be called a warning.
- Tier B Informative pattern Where our best findings sit.
Separates crash-weeks from ordinary weeks in hindsight, but a fire does not reliably precede a crash. A description, not a warning.
- Tier C Worth watching The descriptive tier.
A real, describable sign in the body you can keep an eye on, a weather sign, not a forecast (the framework's tier-1 'monitoring').
Even in hindsight, the timing is at onset
There is one more honest restriction to name. The discrimination question above, looking back, did this signal run differently in the days before a crash than on an ordinary day, has a specific answer on my record. For almost every signal, the difference is on the crash day itself, not the days leading up to it.
A wider sweep (research request OI-038) put eleven of this watch's channels to the same lead-up question: in the days before a crash, do the readings already run different from an ordinary day, or does the difference only arrive on the crash day itself?
On ten of the eleven channels, resting and night heart rate, body-battery floor and morning peak, activity load, sleep duration, and both the sleep-window and all-day stress, the crash-band sits inside the ordinary-day band across all six days before a crash. It only separates on the crash day itself, if at all.
The first pass named this on three signals: the morning body-battery peak, stuck rest-stress, and sleep-stress variability. Across the three, that is eighteen lead-up days in all, and only one of them, a single day on the stuck-stress signal, stood clear of the ordinary band. Every other lead-up day sat inside it; only the crash day separated.
The watch reads the crash, not its approach. It marks the crash as it happens and leaves a recovery signature over the days after; on these channels it does not see the approach coming.
One channel breaks the pattern
On my body the nights start to fray before a crash. Within-sleep awake time runs high as far back as six days out, again five days out, again the day before, and then peaks on the crash day. Sleep fragmentation is the one signal this watch can see that behaves like a genuine multi-day lead-up.
A pre-registered test has since put numbers on it, pooled across the whole record. Across the four days before a crash the size of the run-up is real: the fraying fires on about 73 percent of crash weeks against roughly half of ordinary ones. That four-day gap does not quite clear the bar that tells a real signal from chance (p=0.12). But narrow the window to the last two nights before a crash and it does clear it (a +27-point gap, p=0.04). So the closer to the crash you look, the sharper the fraying gets; at the tightest window it is the one lead-up signal on this record that holds up against chance, though still a modest one.
HA-PG5-precursor, pre-registered and run single-pool across all ~29 crashes (2026-07-22).
This reads the channels this watch records, and the single-pool test settled it about as far as this record can: real, and at the tightest window it clears the bar, but modest, a low hit rate on a rare event. It sharpens the weather report, it does not turn it into a forecast.
The real method behind these numbers (posterior-per-fire and the precision tiers) in the research repo: specificity-tables-spec.md ↗
The lead-up sweep behind the timing section: the wider lead-up sweep (OI-038), research repo ↗