The word list was never the product.

For a long time I believed the product was a spelling app. It had a database of words, a gentle way of correcting mistakes, a schedule that brought old words back at the right moment, and a small screen a child could sit in front of for 5 minutes before school. It worked, in the modest sense that a child came away knowing a word they had not known the day before. And it was the wrong thing to have been proud of.

The part worth keeping was never in the database. What we had built, without quite admitting it, was a written-down theory of how a child learns to spell. Every rule in the product was a sentence in that theory. Track the first attempt, not whether they finished. A small daily dose, not a Monday pile of 18. A hint after the second wrong try, the answer after the third. None of those are interface decisions. They are claims about how a young mind takes in a pattern, and how easily it takes in the wrong one.

This is the same move the protocol itself came from. The duo protocol was never designed up front. It was read back out of a gym app's git history, the patterns that kept recurring, named in hindsight. This ran the same way. The app was the instrument we used to find the theory. I spent six months mistaking the instrument for the discovery.

The artifact is scaffolding. The understanding is the asset. A scaffold is a thing you are meant to be able to take down once the building stands on its own.

So this is an essay about telling those two apart. It is also about a change in the surrounding conditions that means the theory no longer has to live inside the app we found it in.

Here is the thing it took me the whole build to be able to say plainly. The pedagogy is not a collection of teaching tricks. It is a short list of human limits, taken seriously, and then designed around.

Working memory is small. A child can hold only a handful of genuinely new things at once. A school sends home 18 words on a Monday because 18 is the tradition, not because anyone can carry 18. Most never make the journey from the front of the mind into anything durable, and they are gone by midweek. So the right response is not to make 18 words more entertaining. It is to do 5 a day, every day, and let the habit carry them. Daily practice is the product. The small dose is what taking the limit seriously looks like.

Memory fades, and it fades on a curve. A word met again just as it is about to slip away is strengthened far more efficiently than a word crammed the night before a test. Cramming feels productive and mostly is not. Spacing the practice out feels slower and mostly is not. So old words return on their own, brought back by the schedule rather than by a parent's nagging.

The deepest limit is the one that gives the whole project its spine. Practice consolidates whatever you practise, and that includes the error. Each time a child writes a spelling, right or wrong, the act lays down and strengthens a trace. Our own principles put it more plainly than I can. Wrong spellings hard-wire in the brain through repetition, and a hard-wired error is very hard to undo.

That single fact is the entire reason the correction loop exists. A hint after the second wrong try. After the third, the correct answer, shown with encouragement, before the error has a chance to set. From the outside it reads as good manners. It is closer to damage control carried out at the level of memory. The mission was always one sentence. Never let a learner practise the wrong pattern. Everything else is that sentence worked out in detail.

So we had a genuine theory, worked out in some detail. Then we poured it into a fixed mould, and for good reasons.

The shipped app kept its pedagogy in a database. A table of words, sentences written by hand, the tricky clusters marked in advance, the same content handed to every family that opened the door. What adaptivity it had lived entirely in the scheduling. The engine decided when to bring a word back. It never decided what to teach next, and it never wrote a new piece of practice for the specific way a specific child fails. The kid who swaps two letters in a common cluster and the kid who keeps dropping a doubled consonant got the same list and the same hints, because there was only ever one list.

And the one place we most wanted real intelligence, the question of whether this is working and what to do next, we deliberately pushed outside the product. We wrote it down as a principle. Keep the cleverness out of the app, hand the parent their child's data, and let them take it to whatever capable assistant they already had. We even wrote down why. No ongoing costs. A model consulted on every word, for every child, every day, was a meter that would run forever, and we could not justify it. The same documents that named the pedagogy also listed a built-in tutor among the things we would not build, for two reasons. It was expensive, and left to itself it was pedagogically dubious.

Given the conditions at the time, that was the right call, and I would make it again in those conditions. But it is worth naming what it cost. The pedagogy was real, and it was frozen. We knew in principle how to teach the individual child in front of us. We could only afford to ship everyone the average of all of them.

Good decisions carry an expiry date, written into the circumstances that produced them. When the circumstances move, the honest thing is to read the decision again, not to defend the old answer out of loyalty.

There is a seductive misreading of all this, and it fails in exactly the way the pedagogy was built to prevent. The misreading is that models are good now, so point one at a child and let it teach spelling.

A model with no constraint optimises for what looks like success in the moment. Encouragement, momentum, the warm feeling that progress is being made. That is precisely the engagement theatre the pedagogy was built to refuse. Left to its own devices it will happily let a child practise a wrong pattern, because the exchange feels like it is going well. It does not feel the working-memory limit pressing, so it will cheerfully overload. It holds no opinion about the first attempt versus mere completion, because to a fluent generator everything completes.

The theory of learning is what you wrap around that capability so it generates within bounds. Never let a learner practise the wrong pattern stops being a slogan on a wall and becomes a hard edge on what the model may do. It can invent as freely as it likes, but only inside the fence the pedagogy draws. 5 things, not 18. Correct before quick. Surface a word as it fades, not before. Step in before a second wrong attempt becomes a third.

This is the protocol's instinct, stated as plainly as it ever gets. You build the fence at the input rather than relying on a net to catch the mess at the output. The pedagogy is the fence. And the arrival of something powerful enough to genuinely need fencing is exactly what makes the fence worth building well.