So here is the honest version.

What the camera is actually good at

Two things, and it is genuinely good at both.

Identifying what is on the plate. Modern food recognition is strong at naming things. A bowl with grilled chicken, brown rice, black beans, guacamole and corn salsa comes back as five separate items, not one guess at "burrito bowl". That matters more than it sounds: the biggest source of error in food logging has never been arithmetic, it is people not logging at all because typing five entries is tedious.

Being consistent. If you photograph the same lunch on Monday and Thursday, you get roughly the same answer. Consistency is what makes a trend readable, even when the absolute number is off. A log that is uniformly 10% high still shows you the week you ate more than usual.

What it is bad at

Portion size. This is the hard one, and it is hard for physical reasons rather than fixable ones. A photograph is a flat projection of a three-dimensional thing. Depth is inferred, not measured. A bowl of rice photographed from above looks much the same whether it is three centimetres deep or eight.

What it cannot see. Oil absorbed into fried food. Sugar dissolved in a sauce. The butter a dish was finished with. Two visually identical plates of fried rice can differ by a couple of hundred calories depending entirely on things that left no trace in the image.

Dishes it has not met. Recognition quality tracks how well represented a cuisine is in the data behind it. This is why a tracker built around one food culture quietly degrades everywhere else. It is not that it refuses to recognise your food; it is that it confidently returns something adjacent.

What the error actually looks like

The shape of the error follows from the physics above: identification is the strong part, quantification is the weak part, and errors on individual items can be large even when the daily total lands close. Two mistakes in opposite directions often cancel.

That cancellation is the useful insight. It means:

• A single meal's number is a rough estimate. Do not build a decision on it.
• A week's numbers are a usable signal. Errors average out; patterns survive.
• A month's numbers are genuinely informative. This is where tracking earns its keep.

What we do about it in Senumi

Three things, none of them magic.

1. You confirm before anything is logged. The scan produces a list of items with portions, and nothing enters your history until you have looked at it. Untick what is not there, adjust a portion that is obviously off.
2. Estimates are labelled as estimates. Where a number is inferred, the app says so. We would rather be visibly uncertain than quietly wrong.
3. Your correction is what gets logged. Fix a misidentified dish and the corrected version, not the guess, is what enters your history.

The useful way to think about it

A photo estimate is not a kitchen scale, and it is not trying to be. It is closer to a step counter: nobody believes their phone counted exactly 9,193 steps, and nobody needs it to. The number is useful because it is the same kind of wrong every day, which makes the shape of the week real even when the digits are not.

Use it to notice that you eat 600 more calories on days you skip breakfast. Do not use it to decide a dose.

Senumi puts a confirmation screen in the middle instead. That is a deliberate cost, so it is worth saying why we pay it.

What the extra tap buys

It keeps the error correctable. Food recognition is good at naming things and weak at quantifying them. That is not a temporary limitation: a photograph is a flat projection, and depth is inferred rather than measured. So there will be items that come back too large, too small, or occasionally wrong. If those go straight into your history, they are now indistinguishable from the ones that were right. If you see them first, one of them takes two seconds to fix.

It stops the log drifting away from the truth. A tracker is only useful as a record of what actually happened. The moment it becomes a record of what a model thought happened, every conclusion you draw from it is one step removed from your actual eating. The confirmation screen is what keeps those two things attached.

It makes the uncertainty visible. On the confirmation list you can see that the app has guessed "about 1 cup" rather than measured it. That framing carries into how you read the weekly chart, and it should: these are estimates, and an interface that hides the estimating encourages a confidence the numbers do not support.

What it costs

Honesty about the trade: it is slower, and slower means some meals go unlogged. We think that is the right side to err on for a health app, but it is a real cost and not a free win.

We try to make the tap cheap rather than remove it:

• The list arrives pre-ticked, so confirming an accurate scan is one button.
• Untick, do not delete. Removing something the model saw but you did not eat is a checkbox, not a menu.
• A correction is what gets recorded. The list is a draft; your edits are the version that enters your history.

Where the same thinking shows up

Nothing is written to your history until you say so, and nothing is quietly upgraded either. The same rule applies to the parts of the app that reach a server: scanning a lab result or asking the health assistant is something you initiate, not something that happens in the background because you opened the app.

The short version

A number you have checked is worth more than a number you were handed, even when the handed one is usually right. The tap is the difference.

Protein

Protein is the body's building material. Muscle, enzymes, hormones, the structure of skin and hair are all protein. Unlike fat and carbohydrate, the body has no real storage depot for it, which is why intake spread across the day matters more here than for the others.

It is also the most filling of the three per calorie. If you find yourself hungry an hour after eating, the composition of that meal is often a better thing to look at than its size.

Fat

Fat is the densest energy source at roughly nine calories per gram, against four for protein and carbohydrate. It is also structural: cell membranes are built from it, and vitamins A, D, E and K need it present to be absorbed at all.

The useful distinction is not fat versus no fat, it is which fat. Saturated fat is the one health guidance singles out, which is why Senumi tracks it separately rather than folding it into one total.

Carbohydrate

Carbohydrate is the body's most accessible fuel. Starches and sugars both break down to glucose; the difference is speed. A bowl of oats and a glass of juice can carry similar carbohydrate, and behave very differently once eaten.

The carbohydrate in an apple arrives with fibre and water attached; the carbohydrate in a soft drink does not. Senumi tracks total sugar rather than separating out added sugar, so that distinction is one you still have to make yourself.

The two numbers most people should actually watch

Calories are the headline, but for most people two others are more actionable:

1. Protein. Because under-eating it is common, easy to fix, and the fix makes you less hungry rather than more disciplined.
2. Saturated fat and sugar. Because these are the two where health guidance is most specific, and where small habitual changes compound. Senumi gives each its own line on the daily target.

Senumi shows all of these on the daily target, not just the calorie ring. Sugar, saturated fat, dietary cholesterol and trans fat each get their own line, because a day can hit its calorie target and still be a bad day on three of them.

A word on precision

Targets are population averages adjusted for your height, weight, age and activity. They are a starting point, not a prescription. If you have a condition that makes any of this clinically relevant, such as diabetes, kidney disease or a lipid disorder, the numbers that matter for you come from someone who has seen your results, not from an app.