One day, start to finish.

Each line carries where its number came from. Chosen by the assistant is the one to read twice — it means nothing here recognised the food and the model picked a row, so the total is only as good as that pick. A figure whose source cannot be named is not offered at all.

A correction is the thinner case: there is no meal in front of it, so one press is all that stands between the link and insulin. That is why the link expires in half an hour — not for tidiness, but because “he needs a correction” stops being a true sentence about a child fairly fast.

The food still has to be weighed or counted — a meal going somewhere else is not a meal that gets priced more loosely. What changes is where the link goes: say the timing in the same breath and it is emailed to the whole care circle at that time instead of coming back in the chat, so whoever is with him can be the one to press it. “Lunchtime” is not a guess about when people eat — it is read off the meal times you set in Settings, and until you have set them it says so rather than inventing one.

Texts go out from your own number, through Messages on your Mac, to a person or a group you set up on the Texts page — and in the words of a template you wrote, so it reads like you. It is shown before it is queued, it can be changed or cancelled until your Mac picks it up, and the Texts page says Sent from Messages only once Messages has confirmed it. A text carries food and carbs, never insulin.

It will not quietly add a second row that looks like one you already have, and it will not quietly overwrite the old one either. Both are guesses about which food you meant, and the cost of guessing wrong is every future meal priced off the wrong bar.

The review runs by itself, once a night, over the last fourteen days of Loop data — as soon as a Nightscout connection and a timezone are set. There is nothing to remember to do.

Nobody has to ask. Each morning the digest says what the review found, and suggests a change only when the data points to one. When that is a setting, it names the setting, what it is now, the one step to take and the taps in Loop — one step at a time, then a week to see it. When it is waiting on something, it says what, and when it has nothing to say, the subject line says that instead.

What it withheld is reported as carefully as what it found. A withheld finding and no finding are different states, and only one of them means all clear — so a review that stayed quiet about its own silence would be the most misleading thing on the page.

Which source a claim rests on is part of the answer rather than decoration, so the ranking is stated and a source that disagrees is shown rather than dropped. When the review finds something to change, it names the setting, what it is now, the one step to take and the taps in Loop — and nothing changes until you make the change yourself. It is still your decision, with your clinician.

You change settings in Loop, as always. Loop uploads the new profile to Nightscout, the nightly review notices on its own, and from then on it measures the same thing the suggestion was based on, before and after. If you did something Loop cannot show — a habit, a new site location, a talk with your clinician — say so and it goes on the record the same way. It never says a change worked until the measurement does.