How the numbers work

The Catchment exists because church-world numbers are usually borrowed. A statistic appears in a book, then in a conference talk, then in a strategic plan, and by the third stop nobody can say which dataset it came from, which release, or which column. The number survives; its provenance does not.

So the publication has one rule, and it is not negotiable:

No figure appears in print unless it can be reproduced from a named query.

What that means in practice

Every number you read here is declared once, as a query name and a column, in a file called figures.json. A script runs those queries against the Sanctuary warehouse, writes the results to a lock file, and commits it. The prose refers to a figure by name — never by typing the number. At build time the name is replaced with the locked value.

Before an issue publishes, a check re-runs every query and compares the results to the lock. If a value has moved, the build fails and nothing ships until a person has looked at the diff and re-read the sentence the number appears in. That check is the reason this page exists: it is the only thing standing between a small publication and the kind of quiet error that ends one.

Three consequences worth knowing:

A number cannot be typed wrong. There is nothing to typo. If the sentence and the query disagree, the build breaks.

A number cannot drift silently. A new ACS release moves real values. The moment it does, the site stops building, and someone has to decide whether the sentence is still true — not discover it two years later.

You can check us. The receipts table under each issue names the query for every figure. If you have access to the warehouse — or a DuckDB install and the Census summary files — you can run the same query and compare. If your number differs from ours, one of us is wrong, and we would like to know which.

Three kinds of evidence, always labelled

Not every true statement is a recomputed one, and pretending otherwise would be its own kind of dishonesty. So the publication separates three classes and says which it is using:

Recomputed. A number this publication produced from public data with a named query. It carries a receipt. If we cannot recompute it, it is not in this class.

Corroborated. A number someone else published that we have checked against at least one independent source, or against its own underlying table. When sources disagree, we print the disagreement rather than picking the flattering one.

Attributed. A claim from reporting, a denominational body, or a person, which we quote and link but cannot reproduce. "The conference reported X" is attributed. It can be accurate and it is still not ours. Most church-growth numbers in circulation are in this class, and treating them as if they were recomputed is the most common error in the field.

An attribution is not a downgrade. It is an honest label. What is not allowed is printing an attributed claim in the grammar of a measurement.

Limits, published up front

Every issue ends with a "What this doesn't tell you" section. That is not modesty; it is the part of the method most likely to be right when we are wrong. If a reader can name a limit we did not, that is a correction we owe.

Standing limits that apply to almost everything here:

• Geography. ZIP codes (ZCTAs) are postal-statistical areas, not neighborhoods and not parishes. Straight-line distance is not travel distance.
• Vintage. The American Community Survey pools five years. It describes a period, not a moment.
• Modeled sources are labeled as modeled. Where a dataset is itself an estimate — small-area health estimates, vulnerability indices — we say so and do not treat it as a census.
• No outcome data. This publication has neighborhood data, not congregational data. It can tell you who lives nearby. It cannot tell you who will visit, join, stay, or give.

Corrections

If a number here is wrong, it gets fixed in the open: the piece is updated, and a line at the top records what changed, when, and why. Nothing is quietly deleted.

If a whole conclusion is wrong, the correction says that too. A publication that only corrects arithmetic and never corrects judgment is not correcting anything.

Send corrections to [email protected].

Who writes this

One person: Benji Elmgren, a lay volunteer, writing as an independent project. Not a denomination, not a consultancy, not affiliated with any congregation. The tools that fund none of this are here. The data underneath it is described in sources.

AI helps with drafting and with checking drafts against the data. It does not choose the numbers, does not write them, and cannot publish anything. The queries are the queries whether or not a model looked at them first.