Anatomy of one finding.
The other Workings pages explain the parts. This one runs a single real finding through all of them, start to finish: from a thing Willem noticed about his body to a verdict carefully cut down to exactly what the evidence supports. It's a signal that looked, early on, like the strongest on the site, and watching it not survive the whole record is the best way to see how every finding here is held to account.
- 1 The claim
A felt observation becomes a testable question
It started as something Willem noticed about his own body: an intense moment in an otherwise calm day can be enough to trigger a crash. That's a claim you can actually test, but only once you pin down what "an intense moment" means in numbers.
- 2 The pre-registration
Lock the measure and the bar, before looking
The exact measure was fixed in advance: not a day's average stress, but the count of short, sharp stress spikes within a day. And the bar for "this counts" was set before the test ran: a minimum frequency, a minimum gap from chance, a minimum size. Setting that line first is what keeps it honest.
How a question gets locked → - 3 The measurement
From the raw watch data to one number a day
A daily average would bury a five-minute spike under the other 1,435 calm minutes, which is exactly why the earlier average-based test missed it. So the signal is rebuilt from the watch's per-minute stress recordings, with a "spike" defined precisely, producing one honest number per day.
The data dictionary → - 4 The null
Decide what "more than chance" looks like
How many spikes show up in ordinary stretches, just by luck? You answer that by shuffling the data many times to build a picture of the no-signal world, then ask whether the real pre-crash windows stand out from it. Without that comparison, any number looks meaningful.
- 5 The whole-record test
It looked sharp early, then the whole record disagreed
Run over just the early, high-crash years, the pre-crash windows did carry more spikes than chance; it looked like the strongest discrimination in the project. But the honest test is the whole four-year record at once. Judged there, the spike count fired on about half of crash lead-ups and about half of ordinary days: it didn't tell them apart. The early standout was a difference over time, not a verdict.
A number, not a story: the driver ledger → - 6 The stress-test
Why we judge on the whole record, not the flattering slice
This is exactly why the bar is the whole record and not a kind era. Two things sink it. The base rate: even at its early best, when this signal fired a crash followed only about one time in forty; it describes the body, it never warns. And independence: it's the same underlying autonomic signal as a sibling test, so it would count once, not twice. Against the honest bar, it doesn't clear.
A discriminator, not a predictor → - 7 Where it lands
An honest negative: the machine working
So the finding settles where the evidence actually puts it: not found, on the whole record. That isn't a failure of the method; it's the method doing its job. A signal that looked like the project's sharpest, walked through every stage, and came out an honest negative, clearly labelled. Every finding here goes through the same machine, including the one that did hold up.
Seven signals, or one? →
The full pre-registration, test, and result for this finding (H02b), in the research repo: H02b result.md ↗