Skip to content
Download

How we create our content

The whole process, including the role of artificial intelligence — because hiding that would be the only thing here worth a penalty.

Where AI comes in, and where it does not

Artificial intelligence takes part in producing this content. Saying otherwise would be a lie, and leaving it out would be worse: it is precisely what Google’s spam policies describe as content produced at scale with no value of its own. What separates us from that is not the absence of the tool — it is what happens to each piece after the draft exists, and that is what this page describes — in enough detail for you to disagree with the process rather than merely trust it.

AI is used for drafting and for structuring pieces from an editorial plan written by us — the subject, the angle, the references to be discussed and the audience for each piece are decided beforehand, by people. No piece goes live in the shape the draft left it.

What AI never does

  • Write the biblical text. No piece types out the wording of a verse: it declares the reference, and the text is pulled from the edition this site publishes when the page is generated.
  • Settle a doctrinal position. Where the traditions diverge, the note is written by whoever signs the piece, and the policy of not arbitrating is fixed.
  • Sign. Author and reviewer are always two people, each with a name and a profile of their own.
  • Publish without human review of that specific piece.

Human review, piece by piece

Every piece is read in full by someone who did not write it, and the two roles swap on the next piece. Review is not a style pass: it checks whether the piece claims what the cited passage actually supports, whether the commentary describes the text that this edition publishes, and whether the subject required a notice — about disagreement between traditions, or about health — that the draft failed to carry.

See who writes

Reference checking, edition by edition

This is the step that costs the most work and prevents the most errors. Verse numbering differs across the editions we publish: in the Psalms, the Vulgate and Hebrew numbering run a chapter apart through most of the psalter, and in at least thirty books the verse count differs between the Portuguese edition and the Catholic Public Domain Version. Copying a reference from one language to another points, in many cases, at a different verse — one that exists, resolves, and would be published with no warning at all.

The clearest case is Matthew 5: the Catholic Public Domain Version swaps the fourth and fifth beatitudes. A piece about grief that reused the reference would publish the meek under a commentary about mourning, with both chapters holding the same 48 verses — no count would give it away.

So the rule is to resolve every reference in the target language and read the text that comes out before writing the commentary on it. When the commentary leaned on a word the other edition does not use, the commentary changes with it.

The checks that break the build

Part of the control is automated, and deliberately unforgiving: these checks run on every publication and, when they find a problem, the site is not generated. There is no way to publish past them without deleting the check, which would leave a trace.

  • Quotation check

    For every verse quoted in the site interface, it resolves the reference against the published library and compares it word by word with the displayed text. A reference that does not exist in that language’s edition, or text that has drifted too far from what the site actually publishes, breaks the build. This is the check that stops the wording of an unlicensed translation from slipping in.

  • Integrity of the biblical library

    It verifies the SHA-256 signature of each source edition against what was generated, requires the chapters of every book to run contiguously with no gaps, and tests specific passages to confirm the right edition was used — including Psalm numbering, which is what tells one tradition’s edition from the other.

  • Piece registration

    A content file that exists on disk but is not registered in its cluster index never becomes a page, and nothing would flag it: the compiler does not complain about a file nobody imports, the build passes and the sitemap looks complete. This check enforces both directions — a piece with no registration, and a registration with no piece.

  • Valid authorship

    Each cluster’s index is generated, never hand-edited, and the check fails if the generated file differs from what the generator would write. That is what keeps the role alternation from being worked around piece by piece.

  • Links and promised pages

    After the site is generated, it walks the whole output: an internal link pointing at a page that does not exist, an anchor with no target on the same page, an address declared in the sitemap that was never generated, a missing share image — any one of them stops publication. There is also a floor on page count: a sharp drop in the total generated is treated as a failure, not as a result.

  • Indexing policy and attribution

    It verifies that every biblical page shipped with the indexing instruction the policy declares, that no page kept out of the index was declared in the sitemap, and that every chapter page displays the license credit for the translation it publishes — visible attribution is a condition of the license, not a courtesy. The same check fails if any biblical page emits a rating in structured data.

What the checks do not catch

Publishing the list above without this section would give the wrong impression that the process is automatic and closed. It is not, and the limits are known because we found them by getting things wrong.

  • A syntactically correct reference that points at the wrong passage passes every check: it exists, it resolves, and it gets published. Only reading the resolved text catches that — and it has, twice, in pieces whose addresses were perfect.
  • Two editions can publish the same passage making different arguments. A check confirming that a reference points at the same excerpt does not warn you that the other edition says something else there; the piece has to be built on the text that reader will see.
  • No check judges whether a piece is any good. They cover licensing, integrity, registration and links — the rest is human review, which is why it is not optional.

Checks that warn without blocking

Not every automated check should break the build. Two of them compare references across languages and inventory the points where numbering tends to slip, but they only report: the first version of one of them blocked publication and failed two pieces that were correct. A check that rejects correct work is worse than no check, because it teaches everyone to ignore it.

What is generated and what is served

The whole site is static: every page is generated at publication and served as a finished file, with no server deciding anything at the moment you open it. The biblical content is not loaded by script — it is already in the HTML, which means the page behaves the same for a reader, for a screen reader and for a crawler.

Read the editorial policy