Twelve minutes into a session, Long Exhale starts to leave. The tone gets quieter. The cues spread further apart. Three minutes later there is nothing, and nothing announced it.
That is a strange amount of engineering to spend on an ending. It exists because of what the alternatives do.
A guide that stops has to stop at a moment. If that moment lands while you are still awake, you notice the silence, and noticing is the opposite of what the last twelve minutes were for.
Worse, most apps mark the moment. A chime. A gentle "session complete". A screen that brightens to tell you it is finished. Every one of those is a small event aimed at somebody who has, ideally, stopped being able to receive events.
The failure is asymmetric, which is what makes it worth designing around. If the guide fades early and you are still awake, you get quiet a little sooner than you wanted. If it stops abruptly and you are nearly asleep, it can undo the whole session. One of those costs a few minutes. The other costs the night.
Two things move at once, over three minutes:
By the end there is no last cue to notice, because each one was slightly further away and slightly softer than the one before. The session does not finish so much as run out.
On a good night you fall asleep somewhere in the middle of it and never learn how it ended.
The fade normally starts at twelve minutes. It can start earlier, if the app thinks you are already asleep — on Apple Watch, that means your heart rate has settled and your wrist has been still, and both have to agree before anything happens.
That detection is a guess, and it is allowed to be wrong in exactly one direction. Firing early brings the fade forward; it never stops a session. And if it does fire while you are awake, one tap gives you the guide back at the pace it had reached, re-arms the fade five minutes out, and stops the detector second-guessing you for four minutes.
The reason for all that caution is the same asymmetry as before. A detector that is wrong and cannot be corrected drops someone into silence while they are awake — more alert than when they started, and now aware that a piece of software made a decision about their sleep and got it wrong.
There is a hard stop at twenty-five minutes. If you are still awake, the app does not keep going, and it does not ask whether you want another session. Something that will not stop offering itself is a slot machine, and this is meant to be a thing you use at the edge of sleep.
At three in the morning the same guide runs shorter — fourteen minutes, and quieter, because someone is probably asleep next to you and because a twenty-five minute session at 3am is its own source of anxiety.
None of this makes the app better at putting you to sleep. It makes it worse at waking you up, which turns out to be most of the job.