AI & context

The day the AI taught me something about my own system

Where do you put a document that produces nothing but governs everything produced afterwards? The usual test—how fast it changes—is not enough. A test of function replaces it: does this file produce a result, hold on to knowledge, describe a state, or govern future actions? It works on any object the architecture never planned for.

Info

Originally written in French. Translated by AI — the meaning has been preserved, not the prose.

I asked it where to put a translation guide.

It chose the Frame.

I would not have.

The detail might look insignificant. We were talking about one Markdown file in a tree that already holds plenty of them. And yet that choice made me revisit the definition of a central part of the system I have been working on for months.

The unsettling part is that I designed that system myself.

It started with a translation that was too faithful

I am currently working on the English translation of my book about Context.

The first pass was sound. The ideas were there, the structure was respected, and nothing important had gone missing.

But the English still smelt of French.

Some phrasings were grammatically right without being natural. Words came back mechanically from one chapter to the next. Matière almost always turned into material, arbitrage sometimes turned into ruling, and reprise produced wordings nobody would have written straight into English.

The problem was not any single sentence. It was the absence of shared rules to guide the translation of the whole book.

So we started writing a guide: translate the meaning rather than the syntax, keep the personal voice, use natural British English, check how strongly each claim is stated, hold the important terms steady, and reread every text a second time as if the French did not exist.

A simple question came up: where does this guide go?

The book already has its own folder. Each language has one too. The translation scripts, the tracking table and the generated files all live next to the manuscript.

My first reflex would have been to put the guide in the same place.

The AI suggested something else: put it in 00-frame/, at the head of the Context devoted to the book.

The choice looked contradictory to me

In my system, a Context separates several kinds of information.

The Frame carries the mission, the directives and the loading rules. The state shows where the work stands. The sources keep what came in. The notes hold steady what we make of it. The deliverables take in what we produce for a given use and a given reader.

I had always presented the Frame as the stable part.

The translation guide, on the other hand, had just appeared in the middle of the work. It was going to change as we ran into new problems. A few minutes later we were already adding rules to it about reported speech, French realities that are hard to carry over, and the book's conceptual terms.

Why file a document this recent, and this likely to move again, in the zone of stable things?

The suggestion seemed to contradict the architecture.

In fact it was going to help me understand that architecture better.

The guide is not part of the book

To decide where a file goes, I had been looking at what it talked about.

It talked about translating the book. So putting it next to the book felt natural.

The AI had reasoned differently. It had looked at what the file does.

The guide is not meant for the reader. It is not a chapter, not a translation, not an appendix. It is not one more result of the work.

It sets how every future result will have to be produced.

When a new translation starts, the guide has to be read. When a sentence sounds too literal, it gives the criteria for reworking it. When a term drifts from one chapter to the next, it says where to hold that choice steady. When a translation looks finished, it spells out the checks to run before accepting it.

Filing it among the deliverables would have made it a document sitting alongside the others.

Filing it in the Frame made it a rule that applies to every translation that follows.

Once stated, the distinction is simple: the guide does not describe what the Context has produced; it changes how the Context has to produce.

At that point, the AI's choice became coherent.

My definition of the Frame, on the other hand, no longer quite was.

Stable does not mean frozen

Until then I had defined the different zones of a Context mainly by how fast they change.

The Frame barely moves. The state moves often. A source stays fixed. A session grows longer. A draft gets rewritten. A settled decision keeps the record of the choice that was made.

That reading is still useful. It explains why the mission should not be rewritten every time attention shifts, and why what is true right now should not be mixed in with the lasting rules.

But how often something changes is not enough to define what a file is for.

The translation guide was still moving. And yet, as soon as a rule in it was agreed, that rule had to govern every future action of the same kind.

That was when I saw that the Frame was not exactly the zone of what does not change.

It was the zone of what governs the work for the long run.

The difference looks slight. It changes the filing test all the same.

Before, I would have said:

The Frame holds what is defined at the start and hardly ever changes.

Now I would say:

The Frame holds the principles and rules that have to govern future actions for the long run, including the ones discovered and agreed in the middle of the work.

So the Frame is not motionless.

It is constitutional.

It does not change at the pace of daily activity. It changes when the working contract of the Context changes.

Not everything you learn becomes a rule

This new definition opens a risk straight away.

If the Frame can take in a rule discovered mid-work, why not drop into it everything the mission learns?

The translation glossary let me draw the line.

While rereading the English, we picked out several words that are hard to translate: matière, acquis, dette, arbitrage, reprise, cadre, état.

At that stage these are not rules yet. They are things to watch.

Hesitating over the best translation of matière belongs to the work in progress. A suggestion for one chapter is still a hypothesis. A wording that works in one sentence should not automatically become the official translation of that word across the whole book.

But when we decide that the French concept Contexte must always become Context, with the same capital letter and the same formatting, that choice changes status. It no longer just describes what we noticed. It says what every translation from now on will have to apply.

The movement then goes:

terminological hesitation
        ↓
note or state of the work
        ↓ agreed
lasting terminology choice
        ↓
Frame

The Frame can learn.

But it must not turn into a log of everything we learn.

It takes in only what, on examination, has to govern what comes next.

The AI had not decided for me

I could tell this story as the moment an AI understood my system better than I did.

That would make a good hook.

It would also be false.

The AI did not uncover some hidden truth about Context on its own. It applied to a new object the distinctions the system already held: a rule is not a source, a note is not a deliverable, a document that governs production is not itself what gets produced.

It suggested a place and explained why.

My first reaction was that the choice was odd. I put its reasoning up against my definition of the Frame. Then I found that the suggestion held better than my first instinct.

The decision stayed human.

But without the AI's suggestion I would never have run into that weakness in my definition. I would have filed the guide next to the book, the system would have carried on working, and nothing would have forced me to separate what is stable from what governs.

Here, the AI acted as a conceptual mirror.

It did not just hand me an answer. It made me explain why that answer bothered me. Hunting for the flaw in its reasoning, I found the gap in mine.

An architecture has to survive the objects it never planned for

A tree can work for as long as you only meet the objects it was drawn for.

You know where to put a source, a note, a decision and a deliverable because a place was set aside for each.

The trouble starts with the new object.

The translation guide did not exist when I defined the Frame. I had not foreseen that the book would be translated while it was still evolving, nor that those translations would need a set of rules stable enough to hold across several languages and several sessions.

I could have solved the problem by adding an exception: translation guides go in such and such a folder.

But an exception teaches you nothing about the architecture.

The decision gets more interesting when it follows from a test you can use again: does this document produce a result, hold on to knowledge, describe the current state, or govern future actions?

A conceptual architecture becomes robust when it lets you correctly file a new object without having planned for that object in advance.

The test is not whether every known file has a box.

The test is whether the distinctions still help when something the system has never met turns up.

The system handed my own definition back to me

At the start, all I wanted was a better English translation.

By the end, I had moved a guide, built a glossary and changed the way I define the Frame.

I thought it held what did not move.

I now see that it holds what has to govern what comes next, for the long run.

That definition lets the Context learn without confusing what frames it with where it currently stands. A rule can appear late, be argued over, then join the Frame once it becomes binding on future actions. A hesitation can stay elsewhere for as long as it has not earned that status.

So the most interesting detail in this story is not where a file ended up.

It is that, by applying the system to a case it had not foreseen, the AI helped me understand the system itself.

I had built Context so that the AI would not lose the thread of my work.

This time, it was the AI that helped me pick up a thread I had not yet spotted in my own definition.