The danger window that wasn't there.
A test built to find a warning, and it found none.
The most useful thing a watch could ever tell someone with this illness: don't push today, you're not as recovered as you feel. If that warning were real and visible, it would be the one predictive thing on this whole site. So it was put to a real, locked test. Here is what came back, and why an honest "cannot resolve" is the result, not a failure.
- 1 The worry
The one useful thing a watch could tell you
After a crash, the feeling comes back in two or three days, but the body's autonomic load stays raised for closer to two weeks. That gap is where the fear lives: a window where you feel recovered but aren't yet, and a physical push might tip you into a fresh crash. If that danger window were real and visible, it would be the one predictive, actionable thing on this whole site, a reason to hold back. So it was worth a real test, not a hunch.
After the crash: the recovery finding this follows from → - 2 The design
Built to give the idea its best chance
The test wasn't rigged to fail. The exposure wasn't step count but peak cardiac strain relative to my capacity at the time: the cardiac cost of a movement, the measure most likely to catch a real over-reach. Every felt-recovered day in the ten after a crash was compared against a constructed matched baseline: calm-period days at the same fitness level, the same recovery phase, the same recent felt-trajectory. A relapse counted if a new crash or dip arrived within four days. Strongest exposure, tightest matching, aimed squarely at the highest-risk window.
What the watch’s strain signals are → - 3 The lock
Frozen before it ran, provably
As with every claim here: the whole design (the exposure, the window, the matched baseline, the relapse rule, the bar for a verdict) was pre-registered, cold-reviewed in a fresh session, and committed to version control before the test was executed. Then a second fresh session re-ran it from scratch and reproduced every number bit-for-bit. The git history proves the design predates the outcome. There was no room to tune it to a result already seen.
How a question gets locked → - 4 The result
It came back empty
The danger-window pushes relapsed no more often than their matched calm baselines. A third of them were followed by a fresh crash or dip (8 of 24) against roughly 45% for the matched baseline days. If anything the pushes relapsed slightly less, not more: no signal at all in the feared direction. And the error bars are wide enough to swallow the whole question (a standardised difference of −1.03, with a 95% interval running from about −8.6 to +5.2, straddling zero; the result sits near the middle of what pure chance would produce).
- 5 The honest reading
Why this is “cannot resolve,” not “there is no danger window”
Every pre-registered variation gave the same empty result: the window at 7, 10, or 14 days; the relapse counted over 3, 4, or 5; with the matching tightened or loosened. The null is not an artefact of one choice. But it rests on 24 crash windows in one body, and the test was pre-registered as underpowered for anything but a large effect. So a straddling interval reads, by prior agreement, as cannot resolve (the honest floor) never as proof that no danger window exists. The faint negative lean isn't evidence against one either; part of it is a known quirk of how the baseline was built. The finding is held at exactly what it is: a look that could not find a warning, not a guarantee there's nothing to warn about.
Why the watch reads weather, not alarms → - 6 What it earns
A promise kept: there is no crash-risk number to surface
This is the mirror of the recovery-debt question next door, and it lands the same way. The debt doesn't compound; the push doesn't trip a relapse the data can see. Even a purpose-built, best-chance physical test found no warning to raise, so the honest claim that the moment of tipping into a crash simply isn't written into my watch data now rests on a direct test, not only on weak backward-looking signals. And the scope is physical: a crash set off by a hard conversation or a cognitive load is invisible to a strain measure, so this speaks only to the physically-visible kind. Nothing predictive or alarming ships from here, because the whole result is that there is no crash-risk number to put in front of you. On a site that promised a weather report and not an alarm, that's the promise kept.
The other findings we tested and let go →
A null that finds no warning isn't a hole in the site; it's the site working. The one result that would have justified an alarm is the one result the data can't stand behind, and that's reported as plainly as any finding that held. There is no number here that says hold back today, because the honest test for one came up empty.
The locked pre-registration, the test, the fresh-session result review, and the full result, in the research repo: post-crash-exertion-relapse/result.md ↗