Checkable

Who Drew the Line, and What If They Drew It Wrong

The checkable count is a judgement, not a measurement. Closing the part by working out how far the figures move when it is off by one.

Home · Checkable · Caution

Everything in this part rests on somebody having sorted the published statements into two piles. That sorting is an input. The record does not derive it, cannot check it, and would produce figures just as precise if it were wrong.

So the useful question is not whether the line is drawn correctly. It is how much the figures move when it is drawn a statement or two off. For campaigns that pass through several hands, employee activity reports can connect each time record with the project and task that produced it.

Ten published, three depths, five checkable

the whole bar: every content a reader might come away with
the dashed part: contents holding nothing checkable

Three shapes drawn.

Five of the ten published statements counted as checkable in every row. The blind count falls from ten to five to one as the reader meets more.

Computed from the three numbers. A blind content is one holding no checkable statement: that is a property of the page, not a prediction about any reader.

Off by one, on a middling shape

Ten published, five met, five counted as checkable: one blind content out of two hundred and fifty-two. For additional background relevant to the interpretation of online work and information, see Deque axe.

Suppose one of those five does not really survive the test — it names something only the publisher can confirm. The true count is four. The blind figure becomes six of two hundred and fifty-two.

One misclassified statement, and the blind count multiplies by six. In the other direction, a sixth statement turning out to be checkable takes the count from one to nought and closes the question entirely.

Off by one, near the bottom

The same error does much less near the low end of the third column.

Ten published, five met, one checkable: one hundred and twenty-six blind contents. If that one statement does not survive the test, the count becomes two hundred and fifty-two — a doubling. If a second statement turns out to be checkable, it becomes fifty-six.

Doubling, or roughly halving. Large, but not the order-of-magnitude move that the same error produces higher up the column.

Where the sensitivity lives

The pattern is consistent across the record. The blind figure is most sensitive to a misclassification near the threshold, where the count is already small, and least sensitive where it is large.

That is the arithmetic of C(n − c, r). The step factor between consecutive checkable counts, set out in the previous part, sharpens as the count rises and then collapses to nothing at the threshold. Near the top of the column every statement is doing a great deal of work; near the bottom they are nearly interchangeable.

A figure is not trustworthy because it is exact. It is trustworthy when the judgement it rests on is robust, and here the judgement matters most exactly where the figure looks best.

What this means for reading the record's small shares

The blind shares near nought should be read with the least confidence, which is the opposite of how a small number usually reads.

A blind share of one in two hundred and fifty-two is arithmetically exact and rests on a classification where one disputed statement moves it by a factor of six. A blind share of one half rests on a classification where one disputed statement moves it by a factor of two. The reassuring figure is the fragile one.

The record's response is the one it uses throughout: print the input beside the output. Every drawing in this part carries the checkable count as a number on the row, so a reader who disagrees with the classification can substitute their own and recompute. Nothing here asks to be believed.

The figure that is robust to all of it

One result in this part does not move when the classification shifts by one, and it is worth closing on.

If the number of unchecked statements is smaller than the number a reader meets, no blind encounter is possible. That conclusion is a comparison of two integers. It is wrong only if the classification is off by enough to cross the threshold, and when it is not, it holds absolutely — no assumption about which statements are prominent, no assumption about how readers move, no reliance on the uniformity convention.

Every other figure in this part is a count under a convention. That one is a fact about the shape.

A classification that cannot be disputed

One case is immune, and it is worth knowing which.

If nothing published can be checked, the classification is not a judgement. There is no borderline statement to argue about, because there is no statement anybody has claimed is verifiable. Forty-four shapes in the record are of that kind and their figures rest on nothing but the arithmetic.

The same holds at the other end, where everything is claimed to be checkable: the claim is large and it is a single claim, inspectable in one pass. It is the middle of the column — some verifiable, some not — where a reader is being asked to accept a sorting they cannot see, and that is where the record's figures look most informative and are least firm.

What the next part does with all this

The blind count has been treated here as a consequence of the checkable count. The next part takes it as the subject in its own right: how many shapes admit a blind encounter, how large the share gets, where the median sits, and what the figure cannot be used to claim.

The last of those is the important one. A blind share is not a risk, not a probability, and not a prediction about anybody, and the distinction between a count of possibilities and a statement about likelihood is where a record like this most easily goes wrong.