Why one meal proves nothing

You ran the experiment once

Think about what actually happened. You changed one thing — lunch — and you measured the result once, on one afternoon, with everything else about that day left to do whatever it was going to do. Then you read an answer off the chart and filed it away as a fact about the food.

Everything else about that day was not holding still. What you slept, whether you walked anywhere afterwards, what you ate before it, how stressed you were, whether you were coming down with something, how old the sensor was and where it sat on your arm: all of it moves the line, and none of it was written on the chart beside the meal. The rise you saw was the sum of the food and the day. You attributed it to the food.

What happens when the experiment is run properly

Researchers have done the repeat you skipped. Feeding the same person the same meal on different days, with the same brand of sensor on their arm, the glucose response varied from day to day by close to thirty percent.

Sit with that for a moment, because it is the argument in one line. Close to thirty percent is not a rounding error out at the edge of a measurement. It is the size of the difference you were reading a verdict out of. Two meals whose responses sit within thirty percent of each other cannot be told apart by one afternoon each: the ordering you think you saw could easily be the ordering the noise happened to fall in.

There is a hopeful half to the same research, and it is the reason any of this is worth doing at all. Once you have several exposures of the same meal, and you have stayed with one brand of sensor throughout, the ordering of your own meals turns out to hold up well. The noise lives in the individual sitting, not in the pattern underneath it. Repeats are what turn the first into the second.

So a food does not get a number until it has earned one

That finding is not a footnote somewhere in our design documents. It is the design.

• Days, not helpings. In BeeGoodHealth a meal stays marked provisional until you have logged it with sensor coverage on three separate days. Three helpings of the same lunch in one afternoon is one day of evidence, not three, because the variation the gate exists to survive is variation between days.
• Meals the sensor missed do not count. A meal is counted when the glucose record actually covers it: a reading shortly before it to anchor the start, and a reading in the window after it. A meal without both is reported as uncovered rather than quietly averaged in, and it buys no progress toward the three days.
• Until then, the count rather than an average. A provisional meal shows how many times you have logged it and how many more days it is waiting for. An average of one or two responses would look far more settled than it is.
• After that, the spread as well as the middle. Once a meal clears the gate, the spread is reported next to the average rather than hidden behind it, and the number of exposures and the number of days stay on the face of it.

None of that is a trick to make the app feel careful. It is what the day-to-day variation costs you if you want the answer to mean anything: the repeats are not optional, so the app waits for them instead of pretending they happened.

Why the sample size is on the front of the card

Numbers in this category tend to arrive without one. You are shown a score, and no way to tell whether it rests on a dozen careful measurements or on one lunch you happened to log. Leaving the sample size off is what makes a number feel authoritative; putting it on is what makes it possible to argue with.

So the sample size sits beside the number rather than behind a tap, and the arithmetic that produced it is published in the app, in a panel you can open on the free plan as well as on Premium: what counts as the baseline reading, how long the window after the meal is, how the response is worked out, and what the gate is. Meal responses themselves are a Premium feature. The arithmetic behind them is not.

What this costs you

Patience, mostly, and you should hear it here rather than discover it in the app. A dish you eat once a fortnight will sit at provisional for weeks. A meal you describe differently every time will never group with itself, because meals are grouped by what you called them. And a stretch where the sensor was off your arm produces uncovered meals that feel like wasted effort.

We think that is the right trade, and it is a trade rather than a free win. A number that arrives slowly and tells you what it is standing on is worth more to you than a number that arrives at once and cannot be checked. But it is slower, and you deserve to know that before you start.

What this guide will not do

Nothing here sorts your food into a right column and a wrong one. A sample size is not a verdict: it tells you how much a number is standing on, not what to do about it.

And what a pattern in your own meals means for you is a conversation with a clinician who knows you, not with this page. Your history, your medication and a dozen things a website cannot see all bear on it, and we are not in a position to weigh any of them.

What we will do is show you the arithmetic, tell you how many days each number is resting on, and decline to put a score on anything that has not earned one yet.