← notebook

Maria's first data: the borrowed curve runs 35 minutes hot

2026-08-23

Maria's sleep journal went into the app today — 19 days of minute-precision logs, from two months old to two months and eighteen days.

Same routine as her brother: my partner and I write down every wake, every nap, every night waking, and sooner or later it becomes rows in a table. This batch went in through the same validation any hand-logged event has to pass. Within the hour, the habit layer had an opinion.

Her guidance runs on her brother's fitted pressure curve, because a borrowed curve is all a second child can possibly start with. So the first honest question was not does it work but how wrong is it, and in which direction. That is a question most products would rather not ask out loud about their own cold start.

Now I can answer it. Across fourteen comparable days, Maria falls asleep earlier than the borrowed curve predicts at every single nap position — about 45 minutes early for the first nap, then shrinking steadily through the day (37, 35, 33, 22, 12 minutes) — while her bedtime lands about 20 minutes later than predicted.

The satisfying part is that this is not noise, and it is not a bug. It is what the two-process model says should happen to a child whose sleep pressure builds faster than the prior assumes: shorter wake windows, all day long. And the corrections shrinking toward evening, then flipping late at bedtime, fits her age precisely — at two and a half months the circadian contribution is barely switched on, so the evening is expected to behave differently from the day. The prior was wrong in a shape the theory predicts. That is the best kind of wrong.

Two things I want on the record before anyone gets impressed:

The prediction league opened its first round for her today. In a few weeks there will be scored bets, and they will say which method copes best with a clock-emerging infant. The biology predicts the clock-time methods should struggle and the pressure-based ones should hold up.

That is a falsifiable claim, and it is on the record here before the data arrives. If it turns out backwards, this entry stays exactly where it is and I write the correction underneath.