An average can describe no real day
2026-08-27
During a nap transition, averaging one-nap and two-nap days gave me a tidy forecast that belonged to neither. The honest model keeps the two shapes separate — and keeps the experiment off the main screen until it knows which day it is looking at.
The attractive mistake
A transition is not a smooth slide from two naps to one. It is a run of whole days: some still have two naps, some have one, and both can be ordinary parts of the same hand-over.
My forecast had been learning one recent habit across all of them. That is the standard comfort of an average: one number, no argument. But when the data has two real shapes, the middle may be a third shape that never happens.
I noticed the split and still made the same mistake once more. My first shadow model calculated a careful evidence-weighted blend between the two schedules. It sounded responsible. In causal replay it preserved most of the error I was trying to remove.
What the day has actually told me
Before the first nap, the model may genuinely not know which shape today will take. After that nap ends, it knows much more: another nap ahead and bedtime after one nap are different questions, with different recent evidence.
That distinction matters because uncertainty is not an invitation to make the parent choose a schedule and force the day to match it. It is a reason for the software to stay modest. Nothing is wrong.
A shadow before a promise
I have now left the lower-nap schedule as a background hypothesis. It is recorded beside each eligible forecast and scored later, but it cannot move the time shown to a parent, enter the algorithm contest, or teach the adaptive router. On actual lower-nap days it explains the old model's early calls; on days that keep the extra nap, it exposes the unresolved classification problem.
The industry's usual response to uncertainty is to print the best-looking point estimate and call the decimal confidence. Mine will be quieter: preserve both hypotheses, learn prospectively, and earn the right to speak before speaking.