Grounding: Tying an Answer to a Verifiable Source

Grounding means an assertion in the output carries a pointer to a passage somebody can open. Grounding is a property of the output and of the arrangement around the component, never of the component itself. A pointer does not make an assertion more likely to be correct. A pointer makes checking one fast, and checking was what actually cost money.

The difference between a correct answer and a checkable one is easy to slide past, so slow down on it. Almost everybody meets grounding as a quality improvement: attach sources, get better answers. Grounding does nothing of the kind. Nothing about attaching a pointer reaches back inside the component and changes the sentence it produced. The sentence is exactly as right or as wrong as it was going to be. How long it takes a person to find out which is what changes, and at Sumeru Bank Limited, an invented lender, finding out was where the money was going.

What does grounding actually mean, in an arrangement that is running?

Take it out of the bank for a moment. Somebody in a household says the electricity connection has to be transferred before the new tenant moves in. Another member of the household wants to know whether that is true. Said flat, the claim leaves the checking to start from nothing: a look through the agreement, a look through the last set of papers, possibly a telephone call, and at the end of twenty minutes the answer may still be unclear. Now suppose the same sentence arrived with an addition, it is on the fourth sheet of the agreement, second paragraph. The fourth sheet gets opened. Twenty seconds later the answer is known.

GroundingAn assertion carrying a pointer to a passage somebody can open and read for themselves. is the second version. An assertion in the output arrives with a pointerThe reference from a sentence in the output to the passage that is supposed to support it., and the pointer leads to a specific passage rather than to a document or a heading. The parts that did not change between the two versions matter as much as the part that did. The person is exactly as likely to be right in both. Their confidence is the same. The connection either has to be transferred or it does not, and nothing about the way it was said moved that. Grounding operates entirely on the checker's side of the exchange and never on the asserter's side.

Why is grounding a property of the output rather than of the component?

The distinction sounds pedantic and is not. Whether grounding belongs to the component or to the arrangement decides who can be asked to fix a problem. A property of a component is something that can be bought, configured or replaced. A property of an output is something the arrangement around the component produces, and an arrangement is made of retrieval, an instruction, a format and a person, none of which is the component.

Consider what would have to be true for grounding to live inside the component. While producing a sentence, the component would need to know which passage supported it, and it would need that knowledge to be reliable. The component has no such state. At Sumeru Bank Limited three other things happen instead: passages are retrieved and supplied, the 340 word instruction tells the component to attach a reference to the passage it used, and the required format leaves a place for that reference to sit. Change any one of those three and grounding stops happening, without a single thing about the component itself having moved. Grounding is something a deployment arranges, and an arrangement can be put together in a month or quietly stop working in an afternoon.

There is a practical consequence. When somebody describes an arrangement as grounded, the useful questions are all about the arrangement: what is being retrieved, what the instruction says about attaching references, and what the format requires. Asking whether the component supports grounding is asking about the wrong object.

What does it change about the check, and what does it leave alone?

A verificationThe checking a person does on a draft before the output takes effect or gets signed. step asks one question of every assertion: does the supplied material support this? Grounding does not answer that question. It only says where to go and look. The answer can still come back no.

Think about a wedding caterer again, the one from tool calling, but with a different problem. Somebody hands the caterer a note saying two hundred and forty guests have confirmed, and adds, it is on the second sheet of the list. The caterer opens the second sheet and counts. The sheet might say two hundred and forty. It might say two hundred and four. The pointer did not make the note right. The pointer made the counting possible in thirty seconds rather than an afternoon of ringing people. A grounded sentence is no more likely to be correct than an ungrounded one, but the answer can be found in about forty seconds instead of not being found at all.

What is the six minutes of verification actually made of?

At Sumeru Bank Limited, verification of a drafted exception note fell from 9 minutes to 6 across the two measurements. The fall from 9 to 6 is the number the desk quotes. Nobody has asked what the 6 minutes is made of, and the decomposition of it carries the whole argument.

Split it in two. The fixed part is everything the checker does regardless of whether they open anything: reading the draft, checking the required format and the required statements are there, checking that nothing has been asserted which the file does not contain, and handling the decline to answer case where it arises. The fixed part runs to 4.8 minutes. The variable part is opening pointers. A note carries about 6 grounded assertions, opening one takes about 40 seconds, and the desk opens about 30 per cent of the pointers it is handed. Six pointers at 30 per cent is 1.8 pointers, and 1.8 pointers at 40 seconds is 72 seconds, being 1.2 minutes. 4.8 plus 1.2 is the 6.0 minutes the bank recorded, and the second of those two numbers is a choice rather than a fact.

Look at what that split does to the conversation. Before it, the desk had a number: verification takes 6 minutes. Numbers like that get treated as facts about the work. After the split, 4.8 minutes is a fact about the work and 1.2 minutes is the consequence of an opening rateThe share of the pointers handed to a checker that the checker actually follows and reads. nobody ever set. Nobody at the bank decided on 30 per cent. The 30 per cent is simply what happened when busy people were handed more pointers than they had time to follow.

What does a pointer that leads somewhere real but says nothing look like?

Now the part that makes grounding a subject in its own right rather than a paragraph. There is a fault that only exists once grounding exists, and it has no equivalent in an ungrounded arrangement. Call it a miscitationA pointer to a real passage that does not actually support the sentence sitting beside it.: a pointer to a real passage, in the right kind of document and about the right subject, that does not say what the sentence beside it says.

Nothing about the pointer is false. The passage exists. The passage was retrieved. It is genuinely relevant. The document is current. Every property that can be checked about the citation itself comes back clean. The only thing wrong is the relationship between the sentence and the passage, and that relationship is invisible unless somebody opens the passage and reads it against the sentence.

In the second measurement at this bank, 7 of 200 notes carried a statement that was not supported by the material. Of those 7, five carried a pointer to a real passage that did not say what the sentence said, and two carried no pointer at all. Five plus two is seven. Most of what remains wrong at this bank now arrives with a citation attached to it.

Why is a cited sentence checked less carefully than an uncited one?

Because a pointer is read as evidence rather than as an invitation. Nothing more complicated is going on, and the effect is a fact about people rather than about the arrangement.

The same pattern runs through ordinary life. A relative forwards a claim about a scheme and it meets hard resistance. The same relative forwards the same claim with a screenshot attached and the resistance falls away, without anybody opening the screenshot to see whether it says what they think it says. The attachment did not add evidence to the claim. The screenshot added the appearance of evidence, and appearance is what attention responds to.

What does opening every citation cost, and what does it buy?

Both halves of that question have an answer at this bank, and the two answers are straight lines. Opening costs 40 seconds a pointer, so opening all 6 costs 4.0 minutes and any share of them costs that share of 4.0 minutes. Catching goes the same way: the 5 miscited faults sit behind pointers, so opening all of them catches all 5, and opening a third of them catches about a third. The catching line rests on treating the faults as evenly spread across the pointers. The faults are not evenly spread, and the bank never measured how they cluster.

Draw the two lines against each other and the interesting thing is not their slopes. The interesting thing is where the two lines start. The fixed check is going to happen whatever is decided, so the minutes line starts at 4.8. The catching line starts at zero. At full opening, opening accounts for 4.0 of the 8.8 minutes, being 45.5 per cent of the time, and for 5 of the 5 catches, being all of them.

Did the saving survive the opening, once somebody put both in minutes?

Nobody at the bank had run the arithmetic, and it takes about two minutes to run once the decomposition exists.

Before grounding, verification was 9.0 minutes and the fixed check was 4.8 of them, so 4.2 minutes a note was going on tracing assertions to support by searchingLooking for support with no pointer to start from, which is exactly what grounding removes.. Grounding removed the searching, worth 4.2 minutes, and added opening at the desk's 30 per cent rate, worth 1.2 minutes, and 9.0 less 4.2 plus 1.2 is the 6.0 recorded. Put in one number, grounding saved 3 minutes a note. Now open everything: opening all 6 pointers costs 4.0 minutes, so verification becomes 4.8 plus 4.0, being 8.8 minutes.

The 4.2 minutes that grounding removed and the 4.0 minutes that opening everything costs are almost the same number, and nobody had noticed because the two had never been written in the same unit. Opening every pointer leaves verification at 8.8 minutes against the 9.0 it was before grounding existed, a saving of 0.2 minutes a note, being about 12 seconds. Across 3,010 exception notes that is 602 minutes, about 10 hours a month. On time it is a wash. On what those minutes buy it is not a wash at all. The 8.8 minutes catches all 5 miscited faults in every 200 notes. The old 9.0 minutes caught none of them, and none of them could have existed to catch.

Price it, and say plainly what kind of number this is. At the bank's assumed working month of 8,400 minutes and its assumed fully loaded cost of Rs 9,00,000/- a post a year, verification at 6.0 minutes across 3,010 notes runs to 2.15 posts, being about Rs 19,35,000/- a year. At 8.8 minutes it runs to 3.15 posts, being about Rs 28,38,000/-. Moving the desk from 30 per cent opening to 100 per cent therefore costs about Rs 9,03,000/- a year, within a rounding error of one whole post. Every figure in that sentence is arithmetic on the bank's locked and assumed numbers, not a cost anybody at the bank measured.

Where does grounding not apply at all?

Every exception note at this bank has two kinds of sentence in it, and only one kind can ever be grounded. There are statements about what the file says: the declared income figure, the document that was missing, the rule that fired, the date on the statement. For each of those, a passage or a record exists that either supports it or does not. Then there is the judgement: what the file amounts to, whether the gap is the kind that matters here, and what should happen next.

No passage anywhere contains that judgement. Not because the document store is incomplete, and not because retrieval missed it, but because it is a conclusion somebody has to reach about this particular file. There is nothing to point at. Grounding covers the part of a note that describes and never the part that concludes, and the person signing it therefore cannot be replaced by a better arrangement.

The boundary between describing and concluding has a nasty edge to it. If a checking standard says every sentence must carry a pointer, then the judgement sentences fail the standard, and there are only two ways out: drop the judgement out of the note, or attach a pointer to it that does not really support it. The second is how a miscitation gets manufactured deliberately rather than accidentally, and it is worth watching for wherever completeness of citation is being reported as a measure of quality.

How does a desk head, an auditor or anybody signing something actually use this?

Ismail Sheikh, who heads the exception desk at this bank, reads everything above as a rostering question. His people verify every note the component drafts. He now has a knob he did not know he had: the opening rate. Setting it at 100 per cent costs him about one more post a year and returns a complete check. Leaving it at 30 per cent is what happens by default when nobody sets it, and leaving it there costs nothing extra and leaves most of the residue behind citations.

The arithmetic points directly at a middle route, best held as a shape rather than as advice. The 5 miscited faults are spread across 200 notes at about 6 pointers each, being 1,200 pointers, so opening all of them to find 5 faults is expensive per fault: the extra Rs 9,03,000/- a year buys about 53 extra catches a month, being roughly Rs 17,000/- each, all of it arithmetic on the assumed rates rather than a measured cost. Anything that tells a checker which pointers are worth opening changes that price sharply, and the bank never measured how the faults cluster, so nobody there can say what it would change it to. The value of the decomposition is not that it settles the question, but that a question nobody could see is now sitting in front of somebody who can decide it.

Neelima Rao in the risk function, doing the independent validation, uses it differently. Her question is not what the opening rate should be but whether anybody chose it. An arrangement running at 30 per cent because nobody chose is a different thing from an arrangement running at 30 per cent because somebody weighed 53 catches against a post and wrote down why. The number is the same and the accountability is not.

The same shape sits far outside a bank. A household reading a school fee circular, a street vendor handed a wholesale slip, anybody signing a form somebody else prepared: attaching a reference to a claim makes checking it cheap and makes not checking it feel unnecessary. The second effect is stronger than the first unless somebody opens things. A fault can only sit behind a pointer nobody ever opens, so the only useful discipline is to open a few at random.

The last two rows read together give the honest summary of the whole argument. Grounding did not make the arrangement more truthful. Grounding took the same nine minutes of a person's attention and made them buy a complete check instead of a partial one, and it left behind a residue in which most of what is still wrong now arrives wearing a citation.