How Verella decides

Why there is no AI at scan time

A language model asked to rate a food gives a different answer on Tuesday than it gave on Monday, cannot tell you which rule it applied, and will invent a plausible source if you press it. None of that is acceptable when the person asking has Crohn's disease and a peanut allergy.

So Verella scores with a deterministic engine. Every rule is a threshold and a sentence, both written before you scanned anything, both published on this site. The engine is pure arithmetic: same input, same output, forever.

The cost of that choice is honest too. Verella can only know what is on the label and in the open database behind it. When the data is not there, it says insufficient data rather than guessing.

What happens when you scan

1. It looks the barcode up in Open Food Facts, an open database of food products. If the product is missing the basics, calories, protein, carbohydrate and fat, it stops and says so.
2. It works out your targets from your height, weight, age, sex, activity and goal, so a 55 kg person cutting and a 95 kg person bulking do not get measured against the same daily budget.
3. It scores the nutrition against those targets. Protein, fibre, sugar, saturated fat and sodium, each weighed as a share of what your day actually allows.
4. It applies your condition rules. Sixteen of them across six conditions, each one published with its source. Read them all.
5. It reads the ingredient list for the ten flagged ingredients and for your allergens, matching whole words so that "bhaji" never trips the BHA flag.
6. It applies the caps below, which can override everything above.
7. It returns one of four verdicts and every reason that produced it, each with the guidance it came from.

The four caps that beat everything else

Most of a score is a balance. These are not. When one of them applies, good nutrition elsewhere on the label cannot buy the verdict back.

• An allergen you told Verella about

  Caps the verdict at the bottom and is flagged on its own, above everything else. Compound names are matched, so peanut butter is caught as a peanut. This runs for every user whether or not they have anything else set.

• A flagged ingredient a regulator has acted on

  Partially hydrogenated oil, potassium bromate, the six EU warning colours, Red 3, titanium dioxide, brominated vegetable oil or propylparaben. Any one of them forces the lowest verdict whatever the nutrition panel says. The full watchlist.

• A food that does not fit a diet you chose

  If you have told Verella you are vegan, meat is not a deduction. It caps the verdict, because you were never going to buy it.

• Ultra-processed, NOVA group 4

  A softer ceiling than the others. An ultra-processed food can be a perfectly reasonable thing to buy, so it can still reach Fine in moderation. It cannot reach Good choice.

The four verdicts

• Good choice

  Nothing in the label works against you, and several things work for you.

• Fine in moderation

  A reasonable thing to buy. Nothing here needs avoiding, nothing here is remarkable.

• Better occasionally

  Something on this label counts against you often enough that it is worth noticing.

• Best avoided

  Either an ingredient a regulator has acted on, or an allergen you told Verella about, or enough counting against it that nothing else outweighs it.

There is a number underneath, and Verella deliberately does not show it. A number invites you to treat a 72 and a 68 as a meaningful gap when they were produced by different rules firing for different reasons. The verdict plus the reasons is the honest version of the same information.

What Verella will not do

• It will not instruct you. Every condition note is worded as something some people find, because the engine cannot know whether you are in a flare, what your last blood test said, or what your dietitian has already told you. It reports and defers.
• It will not invent a rule. If a concern has no citable source from a named body, it is not in the app. That is why the watchlist is ten ingredients rather than a hundred.
• It will not guess at missing data. A product with no nutrition panel comes back as insufficient data.
• It will not personalise an advertisement. Your conditions and allergens never leave the scoring path.

Download on theApp Store