JESS MOVEFoodLens 360°
Food intelligence
FoodLens returns a range, the source of its evidence and a confidence level — and refuses to collapse that range unless the source is verified. Everything on this page is one real plate, carried through every stage, with every figure computed by the engine rather than written by hand.
One meal, actually read
A weekday lunch: a shop-bought chicken and bacon sandwich, a handful of salted crisps, and an apple.
610to815kcal
A range, because two of the three items were estimated from a photograph. The sandwich is exact — its label was read. The crisps and the apple are not, and a total cannot be more certain than the least certain thing in it.
| Nutrient | Amount | Where it came from |
|---|---|---|
| fat | 28.4glabel | From the sandwich label, plus a reference figure for the crisps by weight. The largest single contributor is the sandwich. |
| saturates | 7.9glabel | The sandwich label carried it. The crisps figure is from a reference table. |
| sugars | 19.1gcalculated | Mostly the apple, worked out from an estimated weight — so this figure moves with how big the apple actually was. |
| salt | 2.6glabel | The sandwich label plus a reference figure for the crisps. This is 43% of a day’s guideline in one lunch, and it is the number worth noticing on this plate. |
| fibre | not measuredunmeasured | Not measured. No label here declared it and a photograph cannot establish it. Shown as unknown rather than as zero, because a zero would be read as a fact about the meal instead of a gap in the evidence. |
The one thing worth noticing: Salt. This lunch is 2.6g, which is 43% of a day’s 6g guideline before dinner. The sandwich carries most of it, and that is a swap rather than a sacrifice.
A photo-calorie app would return a single confident number for this plate. Independent testing presented in 2026 found that leading apps underestimate meal energy by around 250 to 345 calories, with macronutrient error rates between 48% and 66%. The response here is not a better guess — it is a range, a stated basis for every figure, and an unmeasured nutrient left visibly unmeasured.
This example is stored, not generated. Nothing on this page calls a model, spends an allowance or costs anything, and it is the same for everybody who reads it — which also means it is not a cherry-picked run you could not reproduce.
Stage one — capture
Six live checks run on the viewfinder. Each one that passes narrows the cone the estimate will eventually carry — a reference object and a second angle together are worth more than any amount of model confidence.
Each box carries its own confidence, because the interesting part of recognition is not what the model saw — it is how sure it is. Curry sauce at 71% is the weakest link, and it is also the ingredient hiding the most oil.
Two checks failed, so the cone stays at ±18%. Scanning a barcode or adding a side-on photo would tighten it — the app asks rather than guessing.
Stage two — processing
The sequence is visible because each step is a place the answer could go wrong, and you deserve to know which one produced the number you are looking at.
Stage three — the result
A takeaway with a scanned barcode and a confirmed portion scores high. A home-cooked salad photographed in bad light scores low. The number tells you how much to trust the figures beneath it — nothing else.
How much we know about this plate — not a rating of the food.
Reasonable, but the ranges are wide. One correction would narrow them.
Built from four things: the best evidence source available across the plate, how many items were confidently named, how tightly the portion is pinned down, and whether the cooking method — and so the added oil — is known.
Confidence: low — quantity of oil and sauce. A photograph cannot resolve portion size, hidden ingredients or cooking method exactly, so this stays a range.
The macros imply 751 kcal against a stated 690 — +8.8%. Close enough to agree, so both stand.
Salt is high at 1.7g per 100g. The word is printed beside the colour — colour on its own is not an accessible signal.
A user correction outranks everything, including a manufacturer’s label. You know what you put on the plate; the model is guessing.
The Food Intelligence Wheel
One number invites the comparison this product refuses to make, so there isn’t one. Confidence is itself an axis — how well the meal is understood sits beside what is in it, on the same chart.
A permitted framing describes the food. A banned one describes the person eating it. That is the whole distinction, and it is enforced by list rather than by tone of voice.
Allergens
A model can be extremely confident there are no peanuts in a picture and be wrong in a way that puts someone in hospital. So only a complete declaration from a verifiable source may say “absent”. Everything else is unknown, and unknown is displayed as a warning rather than a pass.
Not declared by a source we can verify. We cannot tell from a photograph — check the packet or ask, especially if a reaction would be serious.
The swap ladder
Level one keeps the meal you wanted. “Choose something else” is level five, because a suggestion that ignores what you actually feel like eating is a suggestion nobody takes.
Keeps the same dish, the same sauce and the same rice. Note that the simulated figures carry +8% extra uncertainty — a meal you have not eaten cannot be known as well as one you photographed.
Your pattern
There is no cohort comparison and no percentile against other users, because that is a leaderboard about food. The only useful baseline is the one you set yourself.
Your median evening meal is 647.5 kcal. This one is typical for you — which is a fact about your fortnight, not a verdict about your dinner.
Dinner on Friday and Saturday is doing most of the work. That is a more useful thing to know than a weekly total, because it names a moment you can change.
21 distinct plants this week · new: baked beans, banana, blueberries, coriander, ginger, lentils, oats, red pepper, sesame, spring onion, tomato, turmeric, walnuts, wholemeal bread
A count of distinct plants, not a target you are failing. The empty slots are drawn faintly and carry no red, because an unfinished week is not a bad one.
The boundary
Allergen absence from appearance.
Microbial safety from image.
Disease diagnosis.
Exact calorie count from photo.
Medical treatment advice.
That movement cancels out food.
A single composite health score.
A comparison against another named person.
“A short walk may support your movement target and break up sitting after the meal.”
“Walk for 20 minutes to burn off the burger.”
No calorie, weight or BMI framing reaches a child, in any mode, under any consent setting. bodySurfacePolicy does not consult the switch below 18.
A range with its evidence, every nutrient labelled by how it was arrived at, and nothing invented to fill a gap. Two minutes to an account and the camera is the next screen.
Free to create. 150 AI credits a month for 2 months, then AI features need a plan — everything that is not AI carries on either way. Premium is £5.99 a month and cancels in one click. Under 18 a guardian confirms before the account opens.