Field note
Why capture beats inference
Not because guesses are less accurate. Because a wrong guess cannot be found afterwards, and everything built on it inherits the fault without saying so.
Where this comes from
Two months of running a capture process with an admission gate in front of it: every candidate entry checked against a standard before it was allowed in, and every refusal logged with its reason. The observations about what actually gets refused come from that log. Two months is the age of the gate, not of the system it guards, which has been operating since the end of 2024.
Not published: The log itself, counts of anything, and the contents of accepted or refused entries. What is reported is the kinds of failure and their relative commonness.
The asymmetry
Both kinds of entry can be wrong. That is not the difference.
A captured entry that is wrong is wrong in one place. It has a date, a case and a reason, so when it produces a bad outcome you can walk back to it, look at what you were thinking, and correct it. The damage is bounded by the entry.
An inferred entry behaves differently, and the difference is structural rather than a matter of degree.
The moment it is admitted, it becomes available as a premise. Later inferences use it. Those inferences produce further entries, which are used in turn. None of this is recorded as dependency, because inference does not keep a bill of materials. So when the original turns out to be wrong, you can delete it and nothing else moves. Every entry that rested on it stays exactly where it is, looking locally reasonable, still carrying the fault.
This is why the standard here cannot be an accuracy target. Ninety-five percent accurate inference sounds excellent and is not a usable condition, because you do not know which five percent, and you cannot know what they contaminated. The remaining five percent is not a small residual error. It is an unknown subset of the whole thing.
Why the repair is a rebuild
The natural objection is that this is dramatic. Surely you find the bad entry and take it out.
You can take the entry out. You cannot take out its consequences, because nothing recorded that they were consequences. There is no marking that says this entry exists because of that one. The dependency was real and was never written down.
So the options are to keep a record you now know contains an unknown number of entries derived from something false, or to start again from the material you can verify. Most people, having got that far, choose the first, which is how records like this quietly stop being consulted.
What an admission gate actually refuses
This is the part that cannot be worked out from first principles, and the ordering was the surprise.
Running every candidate entry through a check before admitting it, four kinds of refusal came up. Listed in rough order of how often they occurred.
| Refusal | What it looks like | Why it fails |
|---|---|---|
| The reason restates the call | "Do X, because X is the right thing to do here" | There is nothing to reason from in a new case. It is a fact about one occasion wearing the clothes of a rule |
| Too situational | A rule that amounts to "handle this specific mess the way I handled it" | It will never match another case, so it takes up room and never fires |
| Frustration, not principle | A sharp reaction with no general standard inside it | Real, and correctly discarded. The feeling was information; the entry was not |
| Duplicate | A ruling already held, arrived at again | Not harmful, but it inflates the record and hides that this is a repeat, which is itself a signal |
The restated reason, which is the one to watch for
It was the most common refusal, and it is the most dangerous, because it is the one that feels completed.
You make a call. You write it down. You know that a reason is required, so you write a because-clause. And the clause says the same thing in different words. Move the claim later, because it should not be that early. That reads like a finished entry. It has all the parts. It has a because in it.
It fails the first time a case arrives that the original wording does not cover, which is the only time an entry is worth anything. There is nothing in it to extend from. The call was recorded and the judgment was not.
What a real reason looks like: it refers to something outside the case. A consequence, a mechanism, a thing about how people read or how trust works. If the because-clause can be deleted without losing information, nothing was captured.
Nobody catches this in their own writing in the moment. It is only visible to something checking afterwards against a standard, which is the argument for having a gate at all rather than simply intending to be careful.
Why a smaller record is the stronger one
The instinct is that a bigger record is better and that a gate is a constraint on growth.
The whole value of a body of rulings is that it can be applied without being checked. That property is binary. If any material in there might not be yours, every application requires verification, and verification is the labour the record existed to remove.
So the gate is not a quality filter improving an average. It is what makes the collection a different kind of object from a pile of notes. A modest number of entries that certainly came from real decisions is usable. A large number where some fraction was inferred is a research project.
Frequently asked
Why does capture beat inference?
Because the failure modes are not symmetrical. A captured entry that is wrong is wrong in one place and can be found and fixed. An inferred entry that is wrong becomes a premise for later inferences, and since nothing records which entries rested on it, there is no way to find what it contaminated. The record does not get slightly worse. It becomes unauditable.
Can't you just remove a bad inference?
You can remove the entry. You cannot remove its consequences, because nothing recorded that they were consequences. Everything downstream still looks locally reasonable and still carries the fault. That is what makes the repair a rebuild rather than a correction.
What does an admission gate actually reject?
In practice, four kinds. Entries too situational to generalise. Frustration recorded as though it were a principle. Duplicates of something already held. And, most often and most surprisingly, entries whose stated reason merely restates the correction instead of explaining it.
Why is a restated reason such a common failure?
Because it feels finished. You write the call and then write a because-clause that says the same thing in other words, and it reads like a completed entry. It only fails later, in a case the original wording does not cover, at which point there is nothing in the entry to reason from.
Isn't a smaller record with a gate worse than a bigger one?
No, and this is the point most people get backwards. The value of a record of rulings is that everything in it can be trusted without checking. Admit unverifiable entries and the whole record has to be checked, at which point you are doing the work the record was supposed to save.
Where this sits
The test itself is at captured, not guessed. The integrity condition this describes is at Imprint. The category is defined at what Imprinted AI is.
Related
- Captured, not guessed. The one-question test an entry has to pass.
- What happens when AI guesses your preferences? The same mechanism from the user's side.
- Imprint. Why this is an integrity condition rather than a quality target.
- Imprinted AI vs an AI clone. What happens when a whole record is built by inference.
Keep going
The restated-reason failure is small and it is the reason most attempts at this quietly produce a folder nobody opens again.
The full case is The A.I. Business Manifesto. About 23,000 words, free to read on the page, no gate in front of it. If you would rather have the short version, the same page will send you the three fixes and the PDF.
Read The A.I. Business Manifesto
Free either way. Reading it costs nothing and asks nothing.
Last updated: 28 July 2026