Read these only after you have written your own three sentences. Two of the paragraphs are good, one is not. The question is not which remedy is right — the lesson does not settle that — but what a paragraph has to do to count as a rule rather than a mood.

Model A — nearest the training route

My rule: an expert may give a jury a single figure only after stating, in the same breath, which of the three quantities it is — prior, likelihood or posterior — and the judge repeats the label before the jury retires; judges and expert witnesses are trained to make that labelling mandatory, and an unlabelled figure is struck from the record. This sits nearest the training route, and it protects the swap: Meadow’s “1 in 73 million” would have gone to the jury as “the chance of the evidence if she was innocent”, and the words “the chance she is innocent” could not have been spoken in court by anyone. Where I leave the training route is that I do not train the jury at all — thirty years of commentary have not taught juries statistics, and I am not proposing another thirty; I train the two people who speak, and I give the judge a sentence to say. The price I pay is that a correctly labelled likelihood can still overpower a jury that has no prior to weigh it against, and my rule does nothing to supply the prior; the exclusion route’s court-appointed expert would, and I owe an answer to why a label is enough when the prior for “a mother kills two infants” is itself contested.

What it does: names the rule on the case (what happens to Meadow’s figure, word for word); names the neighbour and the exact point of departure; locates the reason at one step — the swap into the posterior; names the price and admits the debt.

Model B — nearest the exclusion route

My rule: no single-figure probability reaches a jury from a party’s witness; a statistical claim is admitted only through a court-appointed expert who presents the whole count — the prior, the likelihood, and the posterior they produce — in whole people, on one page the jury keeps. This sits nearest the exclusion route, and it protects the prior: Meadow’s figure would not have been excluded, it would have been placed beside the other vanishingly small number, two infant murders in one family, and the jury would have seen that “1 in 73 million” had nothing to be weighed against until that second number was on the table. Where I leave the exclusion route is that I exclude nobody’s evidence, only the form in which it arrives — the DNA match, the epidemiological case, the defendant whose acquittal depends on the base rate all come in, as counts. The price I pay is the court’s expert: someone has to choose the reference class for the prior — all mothers, or families like the Clarks’ — and that choice decides the posterior; I have moved the contested judgment from the witness box to an appointee, and I owe the court a rule for how that appointee is challenged.

What it does: the same four moves, on the other side. Notice that both good paragraphs agree on the arithmetic — once the prior is fixed, the posterior is fixed — and disagree on who fixes the prior and in what form. That is where the case lives.

Model C — a paragraph that does not count

I think the jury was simply innumerate — they were told there was a 1-in-73-million chance that Sally Clark was innocent and could not see that this was wrong, so the answer is better statistical training for everyone involved in a trial. Statistics like this should not be allowed to decide cases, and an expert should never be able to give a number that big. Either way, what happened to Sally Clark must not happen again.

What goes wrong:

The paragraph commits the swap it is meant to prevent: “a 1-in-73-million chance that Sally Clark was innocent” is exactly what Meadow’s figure was not. It was the chance of the evidence given innocence. A rule written by someone who has not kept the two apart cannot keep a jury apart from them either.
The rule is not stated on the case. “Better statistical training for everyone” could have been drafted without the courtroom in view; it says nothing about what would have happened to Meadow’s figure on the day.
No named neighbour, and in fact two neighbours at once: “training for everyone” in the first sentence, “should not be allowed” in the second. A draft that sits on both routes has not yet risked anything.
Both stock failures: training “for everyone”, unspecified against thirty years of recurrence; and an exclusion so broad (“never a number that big”) that it loses the DNA match, the causation case, and the defendant whose acquittal depends on the base rate being heard.
No price. “Must not happen again” costs the writer nothing.

The checklist

A paragraph counts as a rule when it:

states the rule on the case — what it does to Meadow’s “1 in 73 million” specifically;
names which step of the calculation it protects — the prior, the likelihood, or the swap into the posterior — and keeps likelihood and posterior apart in its own sentences;
names its nearest neighbour (training or exclusion) and the exact point where it stops being theirs;
names the price it pays, in the writer’s own words — what the training does differently, or which cases the exclusion keeps.

Model C fails all four. Go back to your own paragraph and mark it 1–4.