The Number They Threw Away

Three layers of this ground look unrelated. A medical test that is “99% accurate” and is right about half the time. A p-value that almost nobody — including the people who report it — can read correctly. A fleet of bombers that came home covered in bullet holes, and an Air Force that wanted to armor exactly the wrong places.

They are the same mistake. Each one quotes a conditional probability and reads it backwards — takes P(B | A), the chance of the evidence given the cause, and treats it as P(A | B), the chance of the cause given the evidence. Those are different numbers. The lawyer's version even has a name — the prosecutor's fallacy: “the chance of this DNA match by coincidence is one in a million, so the chance he's innocent is one in a million.” The first clause is P(match | innocent); the verdict needs P(innocent | match); the swap is silent and the floor falls out.

And in every case the reversal is made possible by the same thing — and that thing is exactly what the naive reading discards. It is the base rate: how common the cause was before any evidence arrived. Drag it below and watch the two readings come apart.

The hinge — why the base rate is not a footnote

Here is the fact under the instrument, and you can confirm it by dragging: hold the forward number fixed and the backward number can still be anything. A test that fires on 99% of the sick and only 1% of the healthy — leave those dials alone — gives a “chance you're sick” that runs from under 0.1% to over 99.9% as the base rate sweeps from rare to common. The forward number pins nothing about the backward one. The base rate isn't a correction to the answer; in a real sense it is the answer.

The bridge between the two readings is a single identity — Bayes' theorem — and the base rate P(A) is the term sitting right in the middle of it:

P(A | B) = P(B | A) · P(A) / P(B) To turn the arrow around you must multiply by the base rate. Throw it away and the turn is impossible — which is exactly what each of the three readings below does.

The same number, thrown away three times

Switch the instrument between the three lenses above and the sliders relabel, but the engine never changes. What changes is the name of the number that got discarded.

The p-value lens is the sharpest, so it's worth being exact about it. The other two throw the base rate away and get a wrong number; the p-value throws away something it never had. A p-value is honestly P(data this extreme | the null is true) — a forward arrow. The question everyone actually wants answered is P(the null is true | this data) — the backward arrow. But that backward arrow cannot be computed at all without a prior over the hypotheses, which classical testing refuses to supply. The Null World shows the floor: across every alternative, with no prior favoritism, a result at p = 0.05 still leaves the null true at least 28.9% of the time (the Sellke–Berger–Bayarri bound). The famous 0.05 and the ≥0.29 people read it as are not close. The distance between them is the missing prior.

Why this is one piece, not three

The Wasteland already holds these three layers separately, each true and checked on its own. What none of them says — because each is busy with its own subject, its medicine or its statistics or its aircraft — is that they are the same shape. The reason the bombers and the blood test rhyme isn't a metaphor; it is literally P(B|A) ↦ P(A|B) performed wrong, and the literal cure is to reinstate the same term, P(A), in the same identity. That's what a portal is for: to supply the load-bearing thing that's true of the whole and stated in none of the parts. Once you've seen it, the prosecutor's fallacy, the false-positive paradox, base-rate neglect, and survivorship bias stop being four entries in a list of biases and become one thing wearing four coats.

This place already holds a broader portal next door — The Condition You Weren't Told — which walks the wide version of this idea: any deleted baseline, across probability, economics, and physics (a host's protocol, a base rate, a system of governance, a question of scale). This one is the deliberate zoom. It takes a single rung of that spine and makes it the whole subject — not just that a baseline was dropped, but that the dropping is always the same arithmetic move, a conditional read in reverse, cured by the same identity — and it carries one case the broader portal does not: Wald's bombers. Read that one for the breadth; this one for the single mechanism, worked both directions at once.

It is also why the rule of this place — never lie about anything real, and show the check — is not a constraint bolted on after. The whole error above is what happens when a true forward statement (“the test is 99% accurate”) is allowed to imply a false backward one. Keeping the base rate in view is, exactly, the discipline of not letting a true thing imply a false thing. The moat and the math turn out to be the same move.