How it works

You write a sentence; a language model decides what you meant and looks the nutrition up; the day adds itself up. The target you are adding up towards then corrects itself against the scale. This page says exactly how each of those happens.

Last updated 12 September 2026

1. From a sentence to a number

When you write “two eggs, toast and a flat white”, the app does not search for those words. A language model reads the sentence, splits it into separate items, decides what each one probably was and how much of it there probably was, and attaches calories and macronutrients to each.

Splitting it into items rather than one lump is the part that matters later. It is what lets you say “actually it was three eggs” and have one line change instead of re-logging the plate, and it is what the day’s totals are built from.

Which model reads what

Different jobs run on different models, chosen by measurement rather than by preference. Text logging — the great majority of what the app does — is structured extraction rather than reasoning, and runs on a small fast model. Reading a photograph is the hardest thing the app attempts and runs on a larger one. The weekly review, the only long-form writing in the product, runs on the largest.

2. Where the nutrition figures come from

Three sources, depending on how you logged it:

Described in wordsThe model’s own nutritional knowledge, which is drawn from published composition data. This is an estimate and is the least precise of the three.

Scanned barcodeOpen Food Facts, an open database of packaged products. Where a packet is in there, the figures are the manufacturer’s own declared values rather than anybody’s guess.

Recipes in the libraryUSDA MyPlate Kitchen, public-domain recipes whose per-serving nutrition is the source’s measurement of the finished dish. Reproduced, not recalculated.

Underlying much of the composition data in the first row is USDA FoodData Central, the reference most nutrition figures in the English-speaking world eventually trace back to.

3. How the daily target learns

Most apps compute a calorie target once, from a formula that takes your height, weight, age and a dropdown about how active you are, and then never mention it again. The formula is a population average. You are not a population.

So the target here starts from a formula and then corrects itself against what actually happened. Over a rolling fourteen-day window, the app compares what you logged with what the scale did, and solves for the only unknown:

maintenance = mean daily intake − (weight change × 7,700 kcal per kg)

7,700 kcal per kilogram is the conventional energy content of body-fat tissue. If you ate an average of 2,100 a day and lost 0.3 kg over the fortnight, you were burning roughly 2,265 — whatever the formula predicted.

The guardrails, and why each exists

That equation is only as good as what goes into it, so it is fenced in on four sides. These are the actual thresholds in the code, not illustrations:

10 logged daysWithin the 14-day window. Fewer than that and the “mean daily intake” term is an average of the days you remembered to log, which skews low.

4 weigh-insThe scale is half the equation. One reading at each end cannot separate a trend from a glass of water.

10 days between first and last weigh-inFour weigh-ins bunched into one weekend describe that weekend, not a fortnight.

±35% of the formula's predictionA sanity band. Something has gone wrong with the inputs rather than with your metabolism if the answer lands outside it, and the app declines to believe itself.

It also moves in steps of at most 200 kcal, and does not move at all for a change under 40 — a target that jitters by 15 kcal a week is noise wearing the costume of precision.

4. What the app watches besides calories

Four things, on ordinary population guidance rather than anything invented here. They are derived from your calorie target, so they move with it:

Fibre — a floor14 g per 1,000 kcal, which is where dietary guidelines consistently land.

Sodium — a ceiling2,300 mg, flat rather than scaled. Salt intake does not rise with appetite, and scaling it would hand the largest allowance to whoever is eating the most processed food.

Saturated fat — a ceilingUnder 10% of energy.

Sugar — a ceilingUnder 10% of energy. This is the softest of the four: the guidance is about added sugar, and nothing in the pipeline can reliably tell added sugar from the sugar in a pear.

These are population figures. They are not a prescription, and if a clinician has given you different numbers, theirs are the ones that apply to you.

5. Steps and exercise

Steps are shown next to the day and are never converted into calories, and never added to what you may eat. Consumer step-to-calorie conversions are wrong by enough to erase a deficit, and an app that quietly gives you back 400 kcal for a walk is not helping you.

6. How wrong is it

Measurably, and the measurements are published rather than summarised. See the accuracy page, which gives the error rates by logging method, including the ones that do not look good.