Glossary
Captured, not guessed
It reads like a value statement. It is actually a test, and you can run it on your own system in about four minutes.
Captured, not guessed (provenance standard). An entry about how a person decides qualifies only if it originated in a decision that person actually made, at a locatable moment. Material derived from finished output, writing style, or patterns across similar people is inferred, and is excluded rather than labelled and kept.
The test
Take anything your AI setup appears to know about how you decide. A preference it applies. A standard it holds you to. A pattern it says you follow.
Ask one question about it.
A captured entry answers. There was a Tuesday, a piece of work, a call you made, and the entry exists because that happened. You may not remember it, but the event is locatable and the entry points at it.
An inferred entry cannot answer, because there was no event. It was assembled from things you produced, which is a different act entirely. The system did not watch you decide. It read the results and worked backwards to a person who would plausibly have produced them.
Run the question on five entries. Whatever proportion cannot answer is the proportion of the picture that was invented, and it is usually higher than expected, because inference produces the most agreeable material. Guesses tend to be flattering and coherent. Real rulings are inconsistent, situational, and occasionally embarrassing.
Why the backwards route cannot work
Not slowly, not partially. Structurally.
A finished piece of work shows the option that survived. It does not show the options that were killed, and it does not show why. Both of those are the judgment. What reached the page is the residue left over after the judgment ran.
Run the arrow the other way and you are asking: given this output, what sort of person produces this? That question has an answer, and the answer is a distribution. The result is the most probable author of that artifact, which is a composite. You are not in it, except in the way an average contains everyone.
The deviations are what get lost, and the deviations were the point. Every time you did the unusual thing that worked, you moved away from the most probable move. Reconstruct you from your outputs and you get the version of you with all of that smoothed off, which is a competent professional in your field, doing the standard thing well.
There is a second loss that gets less attention. Inference from output has never seen anything you rejected. Rejection leaves no artifact. The campaign you killed at the second paragraph, the client you turned down, the feature you refused to ship, none of those exist anywhere to be read. So the reconstruction is built entirely from the yes side of your judgment, and half of a standard is not a weaker standard. It is not a standard at all, because a standard is a line, and you cannot locate a line from one side of it.
Why one guess spoils the set
A good inference is indistinguishable from a captured ruling. That is what makes it good. It uses your vocabulary and matches your pattern, because it was derived from your vocabulary and your pattern.
So the moment inferred material enters unlabelled, the audit above stops working, for everything. You can no longer look at an entry and tell what kind it is. Later reasoning then draws on both kinds equally, which means guesses become the basis for further guesses, and the system's picture of you drifts a step at a time toward somebody who was assembled rather than observed.
Recovery from that state means discarding the set, because there is no way to sort it after the fact. Which is why the rule sits at the point of entry, where it is cheap, rather than at the point of use, where it is impossible.
What this does not say
It does not say inference is bad. Inference is doing useful work all over a sensible setup, and the material it produces is real.
| Layer | Inference here |
|---|---|
| What you produced | Fine. Reading your files is not guessing about you |
| What you decided | Mostly fine. The record often shows the choice |
| How you decide | Excluded. Not recoverable from output, and unauditable once mixed in |
| Who is deciding | Excluded, for the same reason, with higher stakes |
The rule is narrow and it is absolute inside its boundary. Guess about the material. Never guess about the ruling.
Frequently asked
What does captured, not guessed mean?
It is a provenance rule for anything a system holds about how you decide. An entry qualifies as captured if it originated in a decision you actually made, at a moment that can be located. It is a guess if it was derived from your finished output, your writing style, or a pattern across other people who resemble you. Both kinds sound like you. Only one of them is you.
Why does provenance matter if the guess is accurate?
Because you cannot tell which guesses are accurate, and neither can the system. A guess that happens to be right and a guess that is wrong arrive identically, in your voice, with the same confidence. Accuracy you cannot verify is indistinguishable from accuracy you do not have, and it fails at the exact moment you needed it, which is when nobody is checking.
How do I audit a system for this?
Pick one thing it claims to know about how you decide, and ask when you decided that. A captured entry has an answer: a date, a piece of work, a call you made. An inferred entry has no answer, because there was no event. Run it on five entries. The proportion that cannot answer is the proportion of the picture that was invented.
Is inference always wrong to use?
Not at all, and this is worth being precise about. Inference is the right tool for the base layer: what you produced, what you appear to work on, what your files contain. It goes wrong specifically when it is used to fill the layer that records how you rule, because rulings are not recoverable from output and the guesses cannot be separated out afterwards.
Can a system mix captured and inferred material safely?
Only if it keeps the two labelled and never lets inferred entries become the basis for further inference. Unlabelled mixing is the failure case, because the ability to audit disappears the moment the two are in the same pile with nothing to tell them apart.
You can check the implementation
Imprint is open source and provenance preserving, which is to say the standard above is enforced in the code rather than asserted in the marketing. That is a stronger claim than the one most personalisation products make, and it is stronger because it is checkable.
The category is defined at What is Imprinted AI.
Related terms
- Imprint. What a set built entirely of captured entries is.
- The Five Levels of Knowing. Exactly which level inference stops being able to reach.
- Imprinted AI vs an AI clone. The commercial product built entirely on the guessing side.
- Imprinted AI vs AI memory. Where the captured material has to live to stay useful.
Where the term comes from
This is one entry in the vocabulary of a longer argument. The full glossary has fifteen terms. The case they belong to runs about 23,000 words, it is free, and there is no email gate on it.
Read The A.I. Business Manifesto
Nothing on this page is for sale. Quote it, argue with it, or pass it on.
Last updated: 28 July 2026