Where attention goes

Attention is the scarce thing in this work. An assistant produces claims faster than one person can check them, and no amount of diligence closes that gap. So attention is placed deliberately: on the claims that need a person, off the claims a machine can take entirely, and no deeper than the occasion actually requires.

What earns your attention

Two things decide it, and they are asked in this order. Is the claim about to carry weight in something that leaves your desk, and what has the claim already earned. A claim that is going nowhere costs nothing to leave alone, whatever grade it holds. A claim that is going out and rests on nothing but the assistant's fluency is the one case that has to be paid for in your own attention.

Six situations, one of which costs you attention. The vertical axis is whether the claim is going anywhere. The horizontal axis is what it has already earned. Only the cell marked your attention takes a person. The cell beside it takes machine time instead, and moving claims from the first cell into the second is most of what the grades are for.

What the machine takes off you

The grade in the middle of the three is the one that protects your attention, and it is worth being exact about why. It exists because the original documents are kept. A regulation, a standard, a piece of official guidance: the actual file, held locally, catalogued, with a checksum so the file can be confirmed as the one it claims to be. A claim then records which document it rests on, and the exact words it rests on are compared against that file by a string search.

One condition makes that grade mean anything. The comparison has to be run separately from the work that produced the claim. An assistant that writes a note and then awards that note the middle grade on the strength of its own reading has certified itself, and the grade looks identical either way. A check is worth something only when it is independent of the thing being checked.

No person is needed anywhere in that comparison. That is the whole point. The top grade costs your attention and cannot be manufactured, so it has to be rationed. The middle grade costs a few seconds of machine time, so a large body of material can sit at it safely and permanently. Building the store of originals is therefore not archival housekeeping. It is what converts a verification problem that scales with your hours into one that scales with machine time, and it leaves your attention for the claims where judgment is the thing being certified.

what the comparison settlesthe words are in the document, and the document is the one it claims to be. Both are mechanical, both are repeatable, and neither depends on the assistant's account of its own reading.

what it does not settlewhether the conclusion drawn from those words holds. That is judgment, it survives every mechanical check, and it is the reason the top grade exists at all.

where it cannot reacha document that has been machine-translated or read out of an image has no original words to match against. Material in that state stays at the floor until a person reads it, whatever else is known about it.

Nothing is checked in advance

The knowledge base is not swept and verified as a batch. A claim is verified when a piece of work is about to rely on it in front of someone else, and that moment is the trigger. Three things follow.

attention lands where the value isthe claims that get checked are the claims someone is about to rely on. Checking proceeds in the order the work actually needs, which no advance sweep can predict.

unchecked marks the moment, not the materialmaterial is collected because it gets used, and much of that use is thinking, orientation and drafting, where a claim does not yet have to hold up under anyone else's eye. The unchecked mark records that the moment demanding more has not arrived.

the checking compoundsa claim verified for one delivery is verified for every later one. The base matures along the path the work takes through it, so the second delivery in a subject asks less of you than the first.

How deep your own reading goes

The second half of the decision is depth, and it is the half that gets answered badly. The obvious answer is to read the whole source. That is the wrong target.

The rule is that a piece of work is prepared to the level you are willing to speak at, after reasonable preparation, and no deeper. A discussion where you are not the one presenting needs a sound general grasp of the material. The page-by-page read of a long source happens when the work becomes an assignment that demands it, and not before. This does not soften the exit check. Every claim that carries weight is still traced before the work leaves, whatever the occasion. What the occasion sets is how far your own reading goes behind those claims.

Separating the two is what keeps the whole thing inside a working week. The tracing is mechanical and its cost follows the number of claims. The reading is yours, and the occasion is what bounds it.

What this gives up

Two things, both accepted deliberately.

Work can leave carrying claims nobody has verified. They leave marked at the grade they hold, so a reader can tell which claims were checked against a source and which were not. The alternatives are checking everything, which does not fit in a working week, or checking nothing and hoping, which is where the problem started.

The second is that unverified material stays in the base rather than being cleared out. Nothing sweeps it, and nothing is owed beyond the mark it carries. The risk that leaves open is a reader lifting a claim out of the base without looking at its grade, which is why the grade is attached to the claim itself instead of living in a separate register.

The three grades and the two checks are described on the trust page. The procedure for the crossing itself is the audit recipe.