Decisions.

The framework

1. 01

   State the prior as a percentage, and date it.

   You cannot revise a belief you have never stated. Before the pilot, the trial week, the hire, write the claim as a sentence that could turn out true or false, put a percentage beside it and today's date beside that: this pilot converts, twenty per cent; this hire works out, sixty. The number feels presumptuous, and that is its value — a percentage can be moved, checked and argued with, while a good feeling can only be defended or abandoned. Start from the base rate wherever you have one: how often has this sort of thing actually worked out, for you. Watch for the retrofitted prior, the "I always thought it was about fifty-fifty" that appears only once the result is known.

2. 02

   Ask how likely the signal was if you are wrong.

   Evidence should move a belief in proportion to how differently the world would look under each answer, so make two estimates every time something lands: how likely was I to see this if the claim is true, and how likely if it is false. Suppose one pilot in ten converts, and your buyer books a second workshop — a signal that shows up in half the pilots that convert and a fifth of those that don't. That is strong evidence, it more than doubles your odds, and it still leaves you at roughly one in five. The anecdote everyone repeats in the meeting is striking rather than diagnostic: it would have been told just as readily if you were wrong.

3. 03

   Revise in small steps, and count each fact once.

   Two failure modes bracket good updating. The rock sits at ninety per cent and calls it conviction, so every disconfirming signal becomes noise or a bad month. The weathervane overturns a year's position on one loud data point, so the belief carries no information at all. Tetlock's better forecasters lived between them: many small revisions, a few points at a time, without abandoning the base rate they began from. An ordinary observation should move you slightly, and only evidence you would have been surprised to see is licensed to move you far. And count each observation once — the same fact relayed by three colleagues is one piece of evidence, not three.

4. 04

   A number that never crosses into an action is a hobby.

   Name the levels in advance: below thirty per cent we end the pilot and the engineer goes back to the core product; above sixty we hire the second seller this quarter. Setting them before the evidence arrives is what stops a threshold quietly sliding to meet whatever you have come to believe. It also changes the argument in the room, because you stop debating whether things feel positive and start debating the two estimates that produced the number — a debate that can actually be settled. Watch for the evasion: an update that leaves the calendar, the budget and the roadmap exactly where they were.