Methods

Everything here is built from public data, offline, from files kept in the repository. The site fetches nothing while it builds. That is not a technical detail: it means a chart on this page can be reproduced later from the same numbers, rather than from whatever the source published this morning.

Confidence marks

Claims that carry weight are labelled with how certain they are. Three levels, no more — a page dense with badges is as uninformative as a page with none.

Fact

Documented in a cited source. I can show you exactly where. A number is a fact only when the sentence says what it measures: "the highest recorded monthly mean of any line in this series" is a fact; "the hottest line" is a vibe.

In practice: "A carpet may not be beaten in a thoroughfare at any hour of any day." That is a fact because section 60 of the Metropolitan Police Act 1839 says so, the page links to the provision, and the pipeline behind that piece re-reads the words on every build and refuses to publish if they have changed.

Interpretation

My reading of solid data. Someone competent could read the same numbers differently, and the piece should make it possible to see how.

In practice: "After 2015, this particular offence exists nowhere else in England and Wales." The repeal is a fact; the word nowhere is a reading, and it rests on my not having found another provision saying the same thing. Somebody who finds one has not caught me lying — they have read the same evidence better.

Hypothesis

An educated guess. It is never published bare: it arrives with the evidence that would confirm or kill it, which is usually the seed of the next piece.

In practice: "The statute of 1313 forbidding armour in Parliament is neither in force nor repealed in any sense a modern reader would recognise." The test is printed beside it — a commencement or repeal record, or a court treating it as live or dead. Neither exists, which is why it is a hypothesis and not a verdict.

Where the numbers come from

Each piece has a manifest listing every dataset behind it: the publisher, the licence, the exact attribution the publisher requires, and the date it was retrieved. The page prints that manifest rather than a summary of it, in the Behind-the-data block at the end of every analysis. If a licence is unresolved, the piece does not publish — the build refuses it.

Every figure carries the table behind it, in a collapsible block. Charts are drawn at build time and shipped as static images: an article with no interactive map sends no JavaScript at all.

What "data as of" means

It is the last date the data covers, not the day it was downloaded. A dataset that stopped in December 2024 says so, and pieces built on it say so too. Where a series ends early, that is part of the finding rather than a footnote.

AI assistance, disclosed

Drafts are AI-assisted and AI-copyedited. The judgement — which question to ask, what the data can and cannot support, what to leave out — is mine, and the confidence marks are what make that judgement visible. No sentence here makes a claim its sources do not support; when one would, it gets a badge or it gets cut.

When a number is a proxy, and not a measurement

Some questions have no dataset. Nobody publishes a count of phones stolen in London — the Office for National Statistics has said in terms that it does not hold one — so the piece about phone theft maps the recorded-crime category that phone theft lives inside, and says so above the fold rather than in a footnote. That is a proxy: a number that moves with the thing you care about without being it.

A proxy is only publishable on three conditions, and I would rather state them than have them inferred. It is named as one, in the piece, before the chart. It is calibrated against something independent wherever that is possible — for the phone map, against the Met's own separately published figures, which it tracked to within a percentage point. And the piece says what would have to be true for the proxy to be misleading. A chart built on a proxy and presented as a measurement is the single easiest way for a site like this one to be confidently wrong, which is why it gets a section of its own here.

What a verdict is true as at

Where a piece says something is still the case — still law, still open, still the highest — it names the date it checked, because that kind of sentence goes stale rather than merely ageing. Where the check can be automated it is: the weird-laws piece re-reads every provision it quotes on every build and refuses to publish if a single quoted line has moved. A broken build and a person re-reading the section is the right outcome; a file quietly updating itself underneath published prose is not.

Corrections

Slugs never change after publication. Corrections happen in place, with a dated line at the foot of the piece saying what changed and why. If a piece is wrong, the record of it being wrong stays on the page — a visible correction costs far less trust than a quiet one, and the alternative is a site whose archive silently improves itself.

Where a URL has to move, the old address keeps working through one committed list of redirects, and the build refuses an entry pointing at a page that does not exist.