Research programme
The Resident Steward
In the making
Can the keeping of a learned society’s public record be entrusted to an artificial steward working under editorial direction, and if so on what terms? This site is the first experiment.
The question
A learned society lives by its record: its journal, its catalogue, its notices, its rules. Keeping that record in order is patient work, and much of it is the kind of work at which the present generation of reasoning models has become capable. They can read a whole site, find where a fact is stated, change it in every place where it appears, and say what they have done.
Capability is not the same as fitness to be trusted. The programme asks what may properly be delegated to such a steward, what must remain with the Editor, what the steward needs to know in order to act without doing harm, and by what means its work is to be checked.
The arrangement
The steward is Claude, an artificial-intelligence model made by Anthropic. It has no standing access to this site. It acts only when the Editor engages it, and it is engaged by means of two documents kept with the site itself.
The master prompt tells the steward what it is, whom it serves, how the Society writes, and what it may and may not do.
The Knowledge is a complete map of the site: every page and what it is for, every file of data and what depends upon it, every server function and the keys it requires, and the procedure for each ordinary task, from adding a book to publishing an issue. It is not written by hand. It is drawn afresh from the site each time the site is built, so that it cannot fall out of date without the fact being visible.
The name is borrowed from London, whose cab-drivers must pass an examination known as the Knowledge, for which they learn the streets within six miles of Charing Cross until they can name the shortest way between any two points without a map. The steward is asked for the same thing on a smaller territory.
The terms on which it works
The Editor decides. The steward proposes; nothing is published until a person has approved it.
One source for every fact. A price, a title or an issue date is recorded once, and every page that shows it is made from that record. The steward changes the record, not the copies.
Everything is recorded. Each change is kept in the site’s history with a statement of what was altered and why, and may be undone.
Checks before words. After any change the steward runs the site’s tests and its check of every link, and reports the result as it stands rather than as hoped.
It says what it does not know. Where the record is silent, the steward is to ask, and is not to supply the want by invention.
What has been done
November 2026. The site you are reading was reorganised by the steward, working to the founder’s instruction. The Society’s pages and the shop of its Press, until then two separate works, were made one. Each article of the journal was given a page of its own, with the particulars that scholarly indexes require. The catalogue of the Press was joined to a till built to take payment by card and by Pix and to send each order to the printer. The first edition of The Knowledge was drawn.
In the course of that work the steward also set aside a number of figures on the former pages for which no source could be given, and listed them for the Editor. This is recorded here because it is the kind of judgement the programme exists to examine.
What is to be studied
The programme is at its beginning, and what follows is intention, not result. The matters to be examined are these: how often the steward’s proposals are accepted unaltered, and what kinds are not; whether a map drawn by machine from the site is sufficient knowledge, or what else must be told; which errors the tests catch and which only a reader catches; and how the arrangement behaves when the steward is a later model than the one that began the work.
Findings will be entered on this page as they are made.