A number without its base rate is a story
💎CORE ARGUMENT
Give me a number and I will supply a story for it within a second. Suppose someone says that forty exceptions were cleared automatically yesterday. Good, the feature works. But I have not been told how many were cleared on an ordinary day before the feature existed, and without that the forty means nothing at all. A number becomes evidence only when it is set against how often the thing happens anyway.
The mind does not ask for this. It judges by resemblance: the vivid case looks like success, so it is success. The base rate is dull, belongs to no one and has no anecdote attached, so it is left out even when it is sitting in the next column.
The pilot review with Kestrel Cargo on 29 January is where this became a working habit. We had flattering totals to show. What landed was the length of a dispatcher's list on an ordinary morning, because that was the base rate of his day, and every other figure could be read against it. Their operations lead has a shorter way of putting it: show me Tuesday. An average hides the day somebody actually lived.
Three questions now go on every figure before I repeat it:
- Out of how many? A count needs a denominator.
- Compared with when? Last month, last year, before we touched it.
- How often does this happen with no cause at all?
It breaks down when the reference class is wrong. How often do launches like this slip depends entirely on what counts as like this, and people pick the class that flatters them. A base rate does not settle the matter. It starts an argument about which comparison is fair.
🧱BUILT FROM
🌍WHY IT MATTERS
No figure goes into a deck of mine without its denominator and its ordinary-day value. When someone quotes an average at me I ask for one real day, and I have stopped apologising for it.
☑️NEXT STEPS
- Add an ordinary-day column to the pilot dashboard
- Put the three questions on a card by the desk
- Ask Bao for the pre-launch rate of auto-cleared exceptions
Representativeness and availability as shortcuts that ignore prior frequency; the fast answer arriving before the slow question.
The same omission in planning is 💎 My launch estimates ignore every launch I have seen. The same omission in testing is 💎 A significant result is not the same as a true one.
Judgment by resemblance substitutes how typical does this look for how frequent is this. The prior frequency carries most of the information and none of the vividness, so it is dropped. Restoring it turns an impression back into an estimate.
A base rate describes a crowd, and I am usually deciding about one case I know a great deal about. Ignoring what I know about this launch in favour of the average launch is its own error.
A sponsored issue that still said what I meant, a belief I apply daily, and the meeting where the list-length numbers landed.