This site is researched and drafted with the help of AI tools, and every primary or peer-reviewed source in it is opened and checked by a human being before publication. This page explains exactly how that works, because you should not have to take it on trust.
Why this page exists
The value of this site is a single claim: that when we say an inscription says something, it does, and you can go and look. A site built on that claim cannot be vague about how its text is produced. So here is the whole process, including the parts that are uncomfortable.
Where AI is used
AI language models are used here for three things:
- Mapping the state of a question. Which positions exist on the dating of a text, who holds them, what the disagreement actually turns on, and where to go looking.
- Proposing candidate sources. Which artefact, which journal, which critical edition, which locus. Every one of these arrives as a candidate and nothing more.
- Drafting. Structure, first passes at the case for and the case against, and the mechanical work of turning notes into prose.
Where AI is not used
No factual claim and no citation on this site survives into published text on the strength of a model’s say-so. The verdict on a dossier is never delegated: it is a signed editorial judgement, made by a person, and it carries that person’s name.
The failure mode we are guarding against
Language models produce plausible citations that sometimes do not exist. A tablet number, a journal page, a year of publication, a position attributed to a scholar who never held it — this kind of error comes out of a model with exactly the same fluency and the same air of authority as a correct one.
On most websites that is a manageable defect. Here it would be fatal, for three reasons that compound. The asset of this project is verifiability, and a site whose pitch is “we checked the sources” does not survive a reader who checks and finds nothing. The audience for a documentary dossier is, by definition, trained to check footnotes. And in a field occupied by institutions with reputations to protect and apologists with positions to defend, one invented citation in an unfavourable verdict is ready-made ammunition.
So the whole workflow is built around one rule: the model proposes, a human checks, and the software demands proof that the check happened.
What “checked” means here
A source is recorded as verified only when a person has opened the repository that holds it, located the item at the locus cited, and confirmed that it says what the dossier claims it says. Reading an abstract is not checking. Reading a secondary description of the artefact is not checking. Recognising the reference is not checking.
Three things are then recorded, and all three are required: the URL of the repository that was actually opened, the date of the check, and who performed it. The date and the name are stamped by the software at the moment the box is ticked — they are not typed in, precisely so that they cannot be filled in from memory. Changing the repository URL resets the date, because a new repository is a new check.
A primary or peer-reviewed source that has not been verified this way cannot be published, and no dossier that cites it can be published either. This is enforced by the publication gate, not by discipline.
The provenance block
At the foot of every dossier you will find a provenance block: the author and their credentials, the publication date, the date of last review, how many sources sit in each evidence layer, and the date each one was checked. None of it is written by hand. It is generated from the database, which is why it cannot flatter us. If a source is linked but unverified, the block says so in plain words.
The publishing ceiling
We publish no more than four dossiers per month, with a floor of two verified primary or peer-reviewed sources each. This is a hard number rather than an aspiration, and it exists because the real risk of AI-assisted work is not the tool — it is publishing faster than you can verify. When the backlog pushes past the ceiling, what gets cut is the publishing, never the checking.
Assistance does not turn a six-hour dossier into a forty-minute one. It compresses the drafting and the mapping; it does nothing at all to the part where somebody opens an archive and finds the line. Honestly stated: a dossier that would take four to ten hours takes three to six, and the share of that time spent verifying goes up, not down.
If you find an error
Write to contact@scriptureandrecord.com with the page and the problem. Corrections are published on the corrections page with a date, and the affected page is annotated. We would rather carry a visible list of things we got wrong than a quiet reputation for never having been caught.