What didn't survive.
Showing what we kept is easy. Showing what we threw out is the real test. These are findings that looked good (some of them exciting) and didn't hold up when we pushed on them. They're here on purpose.
-
The recovery-arc story
What we thoughtThe body is on a clean recovery trajectory you could draw as a single rising line.
What happenedThe framing didn't survive the project's own 'describe before you infer' rule and was deleted. One person's data can't support a single tidy trajectory; too much else moved at the same time. See what else could explain it.
-
The first 'recent-era warning'
What we thoughtA four-day exertion signal was the first thing that seemed to flag recent crashes: a headline +17-point result.
What happenedIt turned out to be an artifact of how the baseline was computed. On a clean, lagged baseline the effect vanished. The headline was retired the same week it was found.
-
Sleep efficiency as a warning sign
What we thoughtIf bad sleep helps set up a crash, a watch might catch it coming. I tested that with the obvious number: sleep efficiency, how much of my time in bed I am actually asleep.
What happenedEfficiency was flat as a board: not one of my crashes, and not one of 200 ordinary windows, showed even a five-point drop. But that kills the metric, not the idea. Efficiency is a smooth number that can miss what might matter: a night broken into many short wakings, less deep sleep, a shorter night. Bad sleep as a warning may well still stand; I just haven't measured it sharply enough yet.
-
Recovery in zero days
What we thoughtMeasure how long it takes to recover after a crash.
What happenedThe way 'recovered' was defined, every crash recovered the very next day by construction: a fact about the definition, not the body. The test was thrown out and re-written from scratch.
-
The body changed kind
What we thoughtThe early crash and the recent crash are different animals: a real, confirmed change over time.
What happenedTempting, and maybe true, but with one person and a handful of crashes, the design cannot tell a genuine change from medication, pacing, season, or chance. Demoted from a claim to a described difference. See the driver ledger.
A site that only ever shows you wins is selling something. The discipline that produced these burials is the same one standing behind every finding that's still on the board.
The rejected-findings register, in the research repo: REJECTED.md ↗