How we work

Every piece on this site follows the same shape: a question, a model, something you can drive, and the full working underneath. This page is what that means in practice.

The question

A question earns a piece when it is the kind of thing you would argue about at a bar, and tight enough that a number can settle it.

• It has to be answerable. “Is transit underfunded?” is not a question, it is a mood. “What does one metre of this tunnel cost, and what else costs that?” is a question.
• It has to be unmeasured. If someone has already done the arithmetic properly, we link to them instead of redoing it worse.
• We do not run a piece we already know the answer to. If the model cannot surprise us, it cannot surprise you either.

The model

Every answer is a model, and every model is a pile of assumptions. We would rather show you the pile.

• Real unit costs, real rates, real quantities, each one sourced.
• We use the cheapest source that is good enough, and say which one. A national statistics agency where one exists; a vendor price list where it does not; our own measurement where neither does.
• Where the data runs out, we say so and estimate — and the estimate is labelled as one everywhere it appears, not just in a footnote.
• Assumptions are listed in one place, not scattered through the prose. If you disagree with one, you should be able to find it in a few seconds.

Your turn

The model ships as a thing you can drive, not a picture of a thing we drove.

• The inputs we varied are the inputs you can vary. Change the route, the wage, the scope; the answer moves with you.
• Our headline number is just one setting of the controls. It is not privileged, and the piece says which setting it is.

The receipts

At the bottom of every piece, in full:

• every source, linked, with the date it was accessed;
• every fudge factor and what it is standing in for;
• every place the model could be wrong, written by us before anyone else has to.

How we state uncertainty

A confident-looking chart is a claim. We try not to make claims the data will not carry.

• Ranges, not false precision. A number quoted to four significant figures is usually a lie about how well it is known.
• We would rather publish a wide range than a precise wrong number.
• This is enforced in the furniture, not just the prose. The scatter on the front page will not draw a trend line through the archive unless the slope is actually distinguishable from zero at 95% — when it is not, it says so and draws nothing.

Corrections

We will get things wrong. The interesting question is what happens next.

• A correction is a change to the piece and a line in the next email. Fixing it quietly is not fixing it.
• The original number stays visible, struck through, with the date it changed. Pieces are not silently rewritten.
• Errors in the model are treated the same as errors in the prose. A wrong assumption is a correction, not a revision.

The desks

A piece belongs to exactly one desk, chosen by what the question is about rather than who it affects. Each desk has a colour, and that colour is the same everywhere the desk appears — in the navigation, on a tag, and as a point in the scatter on the front page.

What we don’t do

• We do not take money from anyone we write about.
• We do not publish a number we cannot source or reproduce.
• We do not use a chart to say something the underlying data does not.