Body composition, and the charter rule we wrote against it
Charter rule C6 banned body composition framing at every age. Then we specified a product built on exactly those numbers. Here is the whole argument.
What a photograph actually supports
Food photo analysis supports identification of components, reasonably. Relative proportions, roughly. Preparation method, sometimes. Portion mass, poorly — this is the one that dominates the error, and it is the one a photograph is worst at.
A curry can vary by a factor of three in energy density depending on how it was made, and nothing in the image distinguishes the versions. So FoodLens returns a range, and the range is wide when it should be.
The evidence ladder
A barcode is verified. A restaurant menu item with published data is verified. A user-entered weight is strong. A photograph alone is weak. The range narrows as the evidence improves and never narrows because the interface would look better.
This produces the counter-intuitive result that a barcode-scanned ready meal is analysed with more confidence than a carefully photographed home-cooked salad, which is correct and occasionally annoying.
Meal Intelligence scores the analysis, not the food
The number at the top of a FoodLens result is not a verdict on the meal. There is no composite health score — a single number that ranks food good or bad is the shortest path to a disordered relationship with eating, and the Charter forbids it.
Meal Intelligence measures how much the system knows: evidence source, item coverage, portion certainty, preparation certainty. It answers "how well did we read this plate", and the caption saying so is fixed in code rather than left to a designer.
And never to a child
No energy figure, no macronutrient breakdown and no portion judgement is shown below 18, in any mode, under any consent. What a younger user sees is component identification, plant variety, and allergen flags — which is the genuinely useful part anyway.
The twelve dimensions, and why there is no thirteenth
A FoodLens result carries twelve independent dimensions — component identification, portion estimate, energy range, macronutrient split, plant variety, fibre, preparation method, allergen flags, evidence source, confidence, agreement between estimates, and personal relevance against what the person has told us.
What it does not carry is a thirteenth number that combines them. Every reasonable objection to that decision is about convenience, and every objection to the alternative is about what happens to somebody who starts organising their eating around a score that a piece of software invented.
The Charter forbids the composite, and the test suite asserts that no dimension named anything like a health score exists in the response type. It is the most frequently requested feature we have.
Allergens are the one place we refuse to estimate
Fourteen allergens are flagged under UK labelling rules, and a photograph can suggest the presence of most of them. It can never establish absence.
So the result has three states, not two: present, unknown, and verified-absent. The third requires a barcode or published data — it is never inferred from appearance, however confident the identification looks.
The interface says "cannot confirm" rather than showing a reassuring empty space, because an empty space reads as "no allergens" to somebody scanning quickly, and the consequence of that misreading is not a slightly worse meal.
Swaps, and the uncertainty they add
Simulating a change — swap the rice for salad, halve the sauce — is the most useful thing food photo analysis does and the easiest to get wrong. The most useful thing FoodLens does and the easiest to get wrong, because a simulated meal has all the uncertainty of the original plus the uncertainty of the substitution.
So a swap widens the range rather than narrowing it, which is the opposite of what makes a good screenshot. The comparison shown is between two ranges, and where they overlap the interface says the difference is not distinguishable rather than picking a winner.