Glossary

The Imprinted AI Vocabulary

Fifteen terms. Some are ordinary words being used precisely, and some do not exist anywhere else. Each page is an argument rather than a definition, because a definition you can nod at teaches nothing.

If you want the category before the vocabulary, start at what Imprinted AI is. If you want to know how it differs from the things it gets mistaken for, the comparisons do that in eight pages.

The fifteen

Imprint

Turns on: An integrity condition. A body of rulings, defined less by what it holds than by what it refuses to admit. One guessed entry does not make it slightly worse. It makes the whole set unauditable, because nothing marks which entries came from you.

Judgment layer

Turns on: You cannot not have one. It is the only layer of an AI stack that is never absent. Skip it and the position stays occupied, by the model's default, which is a composite of everyone. The choice was never whether to have one.

Captured, not guessed

Turns on: A test, not a virtue. It sounds like a value. It is actually an audit you can run in one question: for any entry, what happened, and when? An entry that cannot answer was inferred.

Digital DNA

Turns on: Two files, deliberately kept apart. Who you are now and who you have chosen to become are stored separately, on purpose. Blur them and the system cannot tell aspiration from evidence, so it optimises you toward the person your history already produced.

Tacit knowledge in business

Turns on: A negative image. Your documentation is a complete record of the knowledge that was easy enough to say. Which makes it, read the other way, a map of exactly where your expertise is not written down.

The Assembler

Turns on: Defined by the question, not the skill. Assembler and Architect can be equally capable and use the same tools. They are separated by the question they bring: what can this do for me, against what should this become inside my business. The market answers the first question all day.

Mirror Score

Turns on: Stand-in rate, per domain. How well does your AI know you is unanswerable as asked. Ask it at decision level and it becomes measurable: on calls you have already made, and cannot see, how often does it make the one you made?

The Five Levels of Knowing

Turns on: Different evidence, not higher rungs. Knows you covers five different things. The break between the second and the third is not a matter of degree. Files can fill the first two and can never fill the third, because the finished work does not contain the decision that produced it.

Belief propagation

Turns on: Impact radius. The expensive part of a changed belief is never the belief. It is the six conclusions still standing on the old one, which all still look correct and none of which announce themselves.

The Final Twenty

Turns on: The only part that was yours. AI delivers the first eighty percent in seconds. You supply the last twenty out of judgment, and it is the only part of the job that was ever yours. It is also the only part with nowhere to go.

Articulation density

Turns on: The seesaw, and what breaks it. Few words and no guessing normally trade against each other. Say less and the machine invents more; spell everything out and you are writing three pages again. Only one thing collapses that trade, and it is not writing better.

Articulation gap

Turns on: You can count it. The distance between what you meant and what you managed to hand over. It stops being a feeling the moment you count the decisions the machine had to invent, and there is a way to make it show you.

Founder judgment

Turns on: The last unpriced input. Every other input a business runs on has a market price. This one never did, because it could not be separated from the founder to be sold. Separating it is the change, and a price is what follows.

Judgment capital

Turns on: Capital earns while you sleep. That is the whole test, and judgment has always failed it. It produced a return only while you were personally applying it, which makes it labour, however well paid. Capture is the conversion.

Apprenticeship without the decade

Turns on: The years were arrival time. The decade was never instruction time. Routine cases take months. The years go to waiting for rare cases to occur while both people happen to be standing there.

Why these fifteen and not forty

A glossary is the easiest thing on the internet to pad. Add a term, write a paragraph that restates it in longer words, repeat. The result reads as filler to a person and gets discounted by every engine that has learned what filler looks like.

So the bar for inclusion was not "does this word appear in the argument." It was whether the entry carries something that appears on no other page here. A few candidates failed that test and are not on this list.

Reading them in order

They are not sequential, but there is a path through if you want one. Tacit knowledge is the problem. Apprenticeship was the only historical solution to it, and the Final Twenty is what it costs you now that you are working with a machine instead of a person. Captured, not guessed is the test that separates a real fix from a plausible one, an Imprint is what passing that test produces, and belief propagation is what keeps it from quietly going out of date.

Three of them are measurable rather than argued, and those are the ones to run rather than read. Articulation gap has a twenty-minute instrument in it. Mirror Score has a ten-case one. The Assembler has a test you can apply to the last five things you asked AI to do.

The longer argument

These terms come out of a document that runs about 23,000 words. It is free and there is no email gate on it.

Read The A.I. Business Manifesto

Nothing here is for sale. If a distinction is useful, the useful thing to do with it is pass it on.

Last updated: 28 July 2026