How this is made
Scrollwork is made by one person with the help of AI models. That is said here, plainly and on its own page, because the whole argument of this site is that you should be able to see where a claim came from — and a site making that argument cannot be quiet about how it was built.
Accurate as of 22 September 2026
What the models do
They draft. They read source material at a scale one person cannot get through by hand — cross-checking a date against the works that disagree about it, finding where a name is spelled four ways, noticing where two accounts of the same event have been dated differently, and setting both readings down.
They also write most of the code, and much of the prose on this site, including some of this page. Reviewed line by line, but written with help. Every word about Scripture is reviewed, and Scripture itself is never generated.
What they are never allowed to do
- Decide what is true. A model can find a disagreement. It does not get to settle one. Where sources conflict, the conflict is recorded as a conflict and shown to you as one.
- Touch Scripture. No verse on this site is generated, paraphrased, corrected, shortened or tidied. Not to fix spacing, not to drop an awkward word. Every quotation is checked against its source text character for character by a step that fails the build if a single character differs — and that check exists because this rule was broken once and nobody noticed until it was looked for.
- Be the source of a fact. Nothing reaches the canvas because a model said so. Every fact carries where it came from, and one that cannot name a source does not ship.
- Run while you read. No model runs when you load a page. Everything is generated and checked in advance and served as static text. Your reading is not sent anywhere, and nothing here is generating itself as you look at it.
Where a person decides
Contested ground is ruled on by me, by name, and the ruling is written down with the reasoning — including the times I was wrong and something had to be corrected. The record of those decisions is kept as the product is built, not reconstructed afterwards.
The honest version of this is that a model is very good at finding the question and has no business answering it. Chronology is the clearest case: where a date is genuinely disputed, the dispute is the fact, and flattening it into one confident number would be the single most damaging thing this product could do.
What is published
Where every fact came from — each fact is shown with the source it came from, so you can check whether a citation resolves or confirm a date carries the chronology it was recorded under. What Scrollwork makes itself — its research, its curation and the records it writes — is its own and is not published as a dataset. Data drawn from a source whose licence asks for it to be shared is shared, under that licence.
The repository itself is not published, and neither is the pipeline, the application or the checks. An earlier draft of this product promised to publish all of it. That promise was withdrawn on purpose, after asking what a published schema actually does for somebody studying the Bible: nothing. It reads as verification without being any, since a running system can differ from a committed file in any way at all. What is published is what a reader can actually use.
Mistakes
There will be some. The interesting question is not whether a product like this contains errors — it does — but whether it is arranged so that an error can be found and fixed by somebody who is not me. That is what the provenance is for.
If you find one, say so. Corrections are treated as defects, and a correction to the record shows as a correction rather than quietly replacing what was there.