You have the same breakfast most days. Same oats, same bowl, same handful of blueberries. Yesterday your tracker called it one number. Today you described it the same way and it came back forty calories higher, and nothing about the bowl changed.
It's a small thing, and it's easy to shrug at. But the second time it happens you stop believing either number, and a tracker you don't believe is one you stop opening.
Every fresh estimate is a new estimate
This happens whenever a number gets worked out again instead of remembered. A person estimating the same bowl twice will land on two answers. A database search can surface a different entry today than it did yesterday. And an AI that reads your description and estimates from scratch will give you a slightly different answer each time too, because that's how these models work — the same words going in don't guarantee the same number coming out.
None of those is anyone being careless. They're all the same mistake: treating a meal you've already logged as though you'd never mentioned it before.
Being consistent matters more than being exact
Here's the counterintuitive part. For what a food log is actually for, a consistent number that's a little off does less damage than an accurate number that wobbles.
Say your breakfast is logged thirty calories high every single day. That error is constant. It shows up in every day equally, so the comparison between days is still honest. Tuesday really was bigger than Monday if the log says so. And if your target gets checked against your weight trend over a few weeks, a steady bias gets absorbed: the target can be corrected to fit what's actually happening, and it doesn't matter that every breakfast was counted a little rich.
A number that jumps around doesn't have that property. Now Tuesday might look bigger than Monday because it was, or because breakfast got estimated on a generous morning. You can't tell which, and neither can anything that reads the log afterwards. It's the same problem a single weigh-in has: noise you can't separate from the signal you're trying to watch. A steady bias you can live with. Noise just makes every day harder to read.
The fix is to estimate once and remember it
The meal you eat every day should come out as the same number every day, because it's the same meal. That doesn't take a better estimate. It takes remembering the first one.
If your tracker has saved meals or a copy-yesterday button, use them for anything you eat more than once a week. It feels like admin the first time and saves you the rest of the time. And re-logging from a saved entry isn't just quicker than describing it again. It's the only way to get the same answer twice.
It also changes what a correction means. When the number on a saved meal changes, it should be because you learned something — you found the label, the portion was bigger than you thought — and you changed it on purpose. That's a correction. A number that moves because it got guessed again is a different thing, and the problem is you can't tell the two apart by looking at them.
What Graham does with this
This is the part of Graham I was most deliberate about, because an AI that reads your description fresh every time is exactly the thing that produces this wobble.
A food you've taught him is fixed. Read him a label once and he saves it, and from then on he logs that food with those exact values instead of looking it up again or working it out again. The foods you've used most recently ride along in every conversation. For anything older, his instructions are to search your own saved list before he reaches for a database. All of that is an instruction he follows, not something the code enforces on its own, and I'd rather say so than claim more. The rule itself is about as blunt as I could write it: same meal, same macros, every time.
"Same as yesterday" means yesterday's entries. When you say it, he pulls yesterday's actual logs and puts the same items in again, values and all. There's a test for it that fails if he asks you what was in it, or if he goes back to the database for it, because both of those would be starting over.
A meal that comes back gets offered as a saved meal. The second time the same combination turns up, he offers once to save it under a name you'd recognise. After that, my breakfast smoothie logs every item in it, with the numbers it had the first time.
"The usual" works without a question. When the same meal lands at the same point in your day on at least three of the last seven, that's a habit, and saying the usual at breakfast logs it.
One more rule, and I'd argue it's the important one. If you correct a saved food later, the fix applies from then on. It doesn't reach back and rewrite last Tuesday. Every entry keeps the numbers it was logged with, so a past day still says what you believed about it at the time, and your history doesn't shift under you whenever you learn something new.
None of that makes your breakfast's number right. It might be thirty calories high forever. What it makes it is the same, and as far as the trend is concerned that's the part that counts.
— Ryan