What we can and can't tell you

Most allergy scanners don't publish a page like this. That is the reason this one exists. Below is what Coreberry actually reads, what it cannot read at all, and what our own testing does and does not prove.

Where a result comes from

A scan takes one of three routes, and they are not equally strong. The app marks which one it used on the result itself — inferred values are shown dashed, estimated numbers carry a ≈.

  1. You scanned a barcode

    The ingredients and nutrition come from the Open Food Facts database — a real label, transcribed by a contributor, at some earlier date. It can be incomplete, and it can describe an older version of the recipe than the packet in your hand.

  2. You photographed a label

    We transcribe the printed panel. This is the strongest route we have: the claim is about a document that exists and that you are holding. It is still a transcription, and a folded, glossy or partly obscured panel can be read wrong.

  3. You photographed a dish

    There is no label. The model names what the food appears to contain, and the nutrition figures are estimates from the portion it thinks it sees. Treat this route as a well-informed guess, never as a reading.

What we do not cover

These are limits of the method, not a backlog. We do not expect to close them.

Cross-contamination and shared equipment

This is the largest gap and the one that matters most. Traces picked up from a shared production line are not part of a recipe, so they are not printed in an ingredient list — which means no label-reading scanner can see them, ours included. The "may contain traces" warning is voluntary: EU rules require the major allergens to be marked in the ingredient list, but say nothing about what a factory line carries over. A product with no trace warning is not a product that was tested and cleared. It is a product whose manufacturer chose not to print one.

Recipe changes

A manufacturer can reformulate between the moment a database entry was written and the moment you pick up the packet. A barcode result can be right about a product and wrong about the item in your hand.

Restaurant and home-cooked food

A plate has no ingredient list. Whatever we say about it is inference from a photograph, and it cannot account for what happened in the kitchen — the oil a dish was fried in, the board it was cut on, a substitution made that morning.

Ingredients a label doesn't have to name

Processing aids, carriers and blanket terms such as "natural flavourings" or "spices" can hide a specific botanical source. We read what is written; we cannot open up what a phrase stands for.

What we measure — and what that measurement is not

We run an internal benchmark on real photographs: 94 of them — dishes, packaged food and cosmetics — across fifteen vision models, each result scored for how well it describes the product. We use it to choose which model to run and to catch regressions before they ship.

That score is a recognition-quality score. It is not an allergen-detection accuracy rate, and it would be dishonest to present it as one — a number like "88% accurate" reads as a guarantee about your allergen, and it would not be one.

We have not put the app through an independent clinical validation, and we are not aware of any allergen-scanning app that has. We would rather say that plainly than publish a figure that sounds like a promise.

How to use this app safely

Related

Last reviewed 2 August 2026