Sources and method
How we test the numbers, and where our information comes from.
How we test accuracy
Estimates are only useful if you know how they were made and checked. Here is our method in short. We do not publish a single headline accuracy number, because it depends on the food, the photo and what is on the label.
The 30-item test set
We keep a fixed set of 30 real foods and re-run it whenever the AI model or prompt changes, and again on a schedule.
- 10 packaged items where the label matters, such as flavored and unsweetened yogurt, protein and snack bars, plant milks, bread and tortillas.
- 10 home plates and simple meals, such as eggs and toast, grain bowls, soup, smoothies and snacks.
- 7 restaurant and takeout items from chains that publish nutrition information.
- 3 tricky items, such as products that look one way and label another.
Ground truth
Packaged items use the printed label times the servings in the photo. Home items are weighed on a kitchen scale and added up from USDA FoodData Central. Restaurant items use the chain's own published nutrition calculator for the exact order.
What we measure
- The typical (median) gap between the estimate and the truth for carbs and sugar, which is what people with PCOS track most. For very small amounts we count the gap in grams instead of in percent.
- Gross misses, meaning an estimate more than 50 percent off on carbs or sugar.
- For packaged items, whether the label or barcode was read correctly.
Rules we hold ourselves to
- A change that makes the typical gap worse, adds gross misses or lowers label reading does not ship.
- We report misses as well as hits.
Label first, and an edit step
- Label first. When you scan a barcode or a nutrition label, the app uses the label's own numbers. The AI estimate is for foods that have no label.
- Everything is labeled. Every AI number is marked as an estimate and each item shows how confident the estimate is.
- Edit before you save. Every number, name and portion can be changed. Missing values such as added sugar are shown as unknown, never as zero.
- Plain rules. The "Plate balance" labels and chips come from simple, published thresholds applied to the numbers you saved. They describe a meal. They are not a health judgment, and the thresholds are our own defaults, not clinical guidance.
Sources we cite
Educational lessons in the app cite their sources by number, with a link, and show the date they were last updated. These are the sources we draw on. Being cited does not mean any of these organizations has reviewed or endorses Cyster Plate.
- World Health Organization: Polycystic ovary syndrome fact sheet
- International Evidence-based Guideline for the Assessment and Management of Polycystic Ovary Syndrome, 2023 (Monash University and partners)
- NHS: Polycystic ovary syndrome (PCOS)
- American College of Obstetricians and Gynecologists (ACOG): Polycystic ovary syndrome
These pages are general information. For anything about you, talk to your doctor.
Recipe sources
Recipes come from US federal nutrition programs. We take the facts (ingredients, amounts, method), rewrite the wording ourselves, and recalculate the nutrition from standard food data. We use no source photos or agency logos. Each recipe shows its credit.
- USDA MyPlate Kitchen (U.S. Department of Agriculture)
- VA Healthy Teaching Kitchen (U.S. Department of Veterans Affairs), VA-authored recipes only
- Keep the Beat (National Heart, Lung, and Blood Institute, NIH)
- Tasty Recipes for People with Diabetes and Their Families (National Diabetes Education Program, NIH and CDC)
Credit wording on recipes:
- Adapted recipes: "Adapted from [collection], [agency]. Recipe source: [partner, if listed]. Nutrition recalculated by Cyster Plate. Not endorsed by [agency]."
- Recipes shown as published from NHLBI: "Recipe courtesy of Keep the Beat™, National Heart, Lung, and Blood Institute."
Recipe tags such as high-fiber or high-protein are computed from the recalculated numbers. Dietary tags describe ingredients only, so check labels for allergens.
Food data
- USDA FoodData Central for generic foods and branded products.
- Open Food Facts for packaged foods and barcodes in our food index. It is a community database available under the Open Database License, and entries can be wrong, so the label scan is one tap away.
Not medical advice
Cyster Plate is a journal and educational tool. It does not make medical judgments about you and it is not a substitute for care from your doctor. Patterns shown in the app compare your own logged days and are not medical findings.