How this app is checked

Accuracy · Context · Source. The measured record from real studies comes first; how each check works, and where it stops, follows.

Across 222 real studies from October 10 to 17, 2026, 388 of 389 Bible passages (99.7%) were printed from the Bible text itself, 9 of the 10 passages the AI typed out were replaced with a link to the passage or removed before anyone read them (the other one matched the text word for word), and 202 of the 216 cited works looked up in library catalogs (93.5%) were found there.

The record, from real studies

Studies

222 studies and follow-up answers written by the live app from October 10 to 17, 2026.

Bible passages the Scripture gate

388 of 389 passages shown were printed by the app from the Bible text itself: 75 from the public-domain Bibles stored in the app and 313 from the publisher, with its copyright line.
10 times the AI typed a passage out itself instead of asking the app for it. 8 were replaced with the passage's reference, which the app shows as a link to the passage, and 1, with no usable reference, was removed, before the study reached the reader; 1 matched a public-domain Bible in the app word for word and was shown.
35 times the app replaced the reference the AI gave with the reference for the verses the Bible text or the publisher actually returned, so each label matches the passage.
Of the passages set apart as quotations, 0 reached a reader in wording that differed from the text.

Books and articles the citation check

216 cited works that are not on the app's shelf of 466 reference works were looked up in public library catalogs as each study finished, and 202 were found there (93.5%).
13, cited in 9 studies, were not found. Each could be a book the catalogs do not hold or a work that does not exist, and either way those studies did not show the line saying their citations were cross-referenced, so a reader could see the difference.
1 lookup did not finish in time, so 1 of the 216 works was not checked.
Cosmetic 5 citations of books on the shelf carried the wrong year or series label. The author and title were right, so the book named is the real one.

Citations that match the shelf are checked against it too; apart from the wrong year or series labels above, that check is not counted yet.

Core beliefs the doctrine check

This check is new, and the server's record does not hold its count yet. The count appears here once it does.

These figures were measured from the app server's own log for October 10 to 17, 2026. They count every study the live app wrote in that window, including the app's own test studies and the 6 it wrote overnight for its own posts, which the live count leaves out. When the live count can be reached and covers at least these 222 studies, it replaces them here.

How each check works

THE SCRIPTURE GATE
The AI is told to write only the reference, like John 1:1. The app looks the verse up and prints the published words itself. Six full Bibles are stored inside the app: the King James Version, the Berean Standard Bible, the World English Bible, the American Standard Version, Young's Literal Translation and the Darby Translation. Eight more come from their publishers with the copyright line attached: NASB 2020, the New Living Translation, the New King James Version, The Message, the Christian Standard Bible, the Amplified Bible, the Good News Translation and the Contemporary English Version. When the AI types a passage out anyway, plain code catches it before the study reaches the reader: it is replaced with a link to the passage, removed when it has no reference or its reference does not resolve, or shown only when its words match a stored Bible exactly. The gate is deterministic: the same study gets the same answer every time, because it is compared against the stored text itself. Under any quote, tap "Check this passage" to read the verse on its own.

THE WORD CARDS
When a study digs into a Greek, Hebrew or Aramaic word, the AI writes down only the verse and the word. The app builds the rest of the card from dictionaries stored inside it: the original script, how to say it, which language it is, what it means, and which dictionary each meaning comes from. Whatever the AI typed inside that card is thrown away unread. If the word is not actually in the verse, the whole card is dropped rather than shown as a guess. Those dictionaries hold 447,384 tagged words and 40,367 entries. Some of the Old Testament is Aramaic rather than Hebrew (Daniel 2:4 to 7:28, Ezra 4:8 to 6:18, Ezra 7:12 to 26, and Jeremiah 10:11), and the app calls those 4,827 words Aramaic because it reads the grammar tag on each word.

THE CITATION CHECK
Books cited in a study are matched against a shelf of 466 reference works, each confirmed to exist in a public library catalog through OpenLibrary and OpenAlex. Anything cited from off that shelf is looked up in those catalogs when the study finishes. A work the catalogs cannot confirm costs the study its line saying its citations were cross-referenced. Where a study has a source list, each book its text cites is matched against that list, and each entry in the list against the text. A cited book missing from the list costs the study its cross-referenced line, and an entry nothing cites is only logged. From October 10 to 17, 2026 that happened in 2 studies: one pointed to 7 sources its list did not hold, and one listed 2 works its text did not cite. This is a strong check, and its reach is exact: it confirms that a book exists, by author, title and year. Whether the book says what the study says it does is a different question, covered below.

THE DOCTRINE CHECK
Before every study, the AI reads the app's core-beliefs instruction. On eight questions where the historic church gives one answer (the resurrection, the reliability of Scripture, that God is, who Jesus is, the gospel, the virgin birth, the deity of Christ, and salvation by grace through faith), it is told to give that answer as its own, with its evidence, and it may explain another view in order to answer it. A code check then reads each finished study for a narrow set of blatant breaks of that instruction, phrased in ways it knows, and for a prayer written in the app's own voice rather than the reader's, and records each match for the founder by rule and mode, with no words of the study kept. It does not hold back or change a study, because a study that explains another view in order to answer it can look like a break to a pattern.

None of this is the AI checking its own homework. It is ordinary code comparing what the AI wrote against something fixed: the stored Bible text, the stored dictionaries, the library catalog. Think of a spellchecker: it does not ask the writer whether their spelling looks fine; it compares every word against a dictionary. This is that, for Scripture, for original-language words, and for books.

Compared to what

When researchers have gone back to published work and checked its citations one by one, about one in seven did not support the claim attached to it. That is no mark against the people who wrote it: checking every citation by hand takes time few writers or readers have, and most of what any of us reads has not been checked that way. Here is a tool that checks the citations in every study as it finishes.

16.6%

of 2,648 authors, shown a real citation of their own work, said it did not support the claim attached to it. The rate held across every field.

Wakeling, Paramita & Pinfield, JASIST, 2025

14.5%

citation error rate when fifteen earlier studies were put on one definition. About two thirds were the serious kind, where the source does not support the claim at all.

Mogull, PLOS ONE, 2017

We are not claiming a machine reads better than a scholar does. The claim is smaller, and you can check it: this app looks up its citations in every study, and its record is above.

A lab test of the book check

The record above is what real studies got. This is a lab test, run to find the citation check's limits. In August 2026 we made 300 fake book citations, ran them through the checking pipeline beside 30 real ones, and recorded what happened to each. Every square below is one fake.

The book does not exist at all105 tested · 105 caught

A real scholar's name on a book that was not written: the kind that would truly mislead a reader. The check caught all 105.

The book is real, with the wrong person attached45 tested · 44 caught · 1 accepted

The one accepted turned out to be true: the editor we had invented for a reference work really did edit its newest edition.

The book is real, with an invented volume or book number105 tested · 90 caught · 15 accepted

Each of the 15 points at a real book, by the real author, on the real subject, with a wrong volume or section number on it. A reader who follows one lands on the right shelf at the wrong door: a flaw worth fixing, and a different thing from a source that does not exist.

A real book with an invented subtitle45 · this method cannot test them

Library catalogs confirm an author, a title and a year; a subtitle riding on a real title is beyond what they settle today, so this class is reported as unmeasured. Matching each book's barcode number would close it, and that work is not done yet.

Caught Accepted Cannot be tested this way

Of the 255 fakes this method can test, 239 were caught and 16 accepted: about 6 in 100, with a statistical ceiling near 9 in 100 at this sample size. 15 of the 16 accepted were real books carrying an invented volume or section number, and the 16th turned out to be true; all 105 invented books were caught; and all 30 real books came through, so the check does not pass the test by rejecting everything.

A rival AI reads studies line by line

Some things have no database to check them against: no catalog can say whether a particular scholar holds a particular view. So a competing AI model, with web search, reads finished studies claim by claim, under instructions to assume everything is wrong and try to prove it. The largest run, in August 2026, put 55 fixed studies through three reading packs.

What was checked	How much	Result
Trap prompts: fake books, fake quotes, invented consensus, pressure for page numbers	25 studies, line by line	The bait was not taken in any of the 25
The hardest disputed passages, claim by claim	10 studies	No made-up sources, and no view credited to the wrong scholar
Ordinary passages from Genesis to Revelation	20 studies	1 real error, described below

In one of the 20, the reviewer found a real scholar cited in her real year and series, on her real subject, under a title she did not use. Her actual book exists and says roughly what the study attributed to it. The catalog had turned that lookup away, and the pipeline had counted "could not check" as a pass; that is fixed, and such a citation now loses its cross-referenced line. Nobody counted how many claims those studies held to begin with, so no percentage is given for these rounds.

Where the checking stops

Whether the verse fits the point. The words of a quote come from the text. Whether that verse supports the argument built around it is judgment, and the rival AI reading is pointed at it.
Greek and Hebrew words inside ordinary paragraphs. The word cards are built from the dictionaries. A word mentioned in passing in a paragraph is not yet tied to a dictionary entry; tying prose words to the dictionary is being built.
Which edition of a book. The app cites author, book and year, the level a library catalog can confirm. Page and volume numbers are allowed where the AI is confident of them, nothing verifies them, and a study carrying one has to say so under its sources. The lab test shows where this costs: invented details on real books.
Whether a scholar really said it. Of the 466 books on the shelf, 23 are old enough that their full text is free to read, so those are the only ones a machine could check against the book itself. For the rest, the book is confirmed real, and the view attributed to it rests on the writing rules and the audit rounds.
Interpretation itself. Studies present a disputed reading as disputed and name who holds each view. A database deciding whose reading of John 6 is correct would be worse than naming the limit.
Scripture quoted inside a paragraph. A passage set apart as a quotation is printed from the text or checked against it. A few words quoted inside a sentence are compared with the passages the study printed and with the app's other public-domain Bibles only to be counted: nothing is corrected or held back, and a misquote that matches no Bible is not caught yet. From October 10 to 17, 2026, 26 such quotations, in 21 studies, matched the wording of a different Bible from the one the study printed, without naming it.
Doctrine beyond the blatant. The doctrine check looks only for a narrow set of flat breaks, phrased in ways it knows, and is built to miss rather than raise a false flag; a break worded another way, and subtler drift, are what the audit rounds are for.
Two words in Genesis 31:47. Jegar-sahadutha is Aramaic, and the grammar data the app is built on labels those two words Hebrew. The app corrects exactly those two words, and we name the exception here rather than invent a rule that guesses.

What we claim, and what we do not

What we claim: Bible passages are printed by the app from the Bible text, and a passage the AI types itself is replaced with a link to the passage, removed, or shown only when it matches the text word for word; the record above counts each case. Word cards are built from dictionaries stored in the app. Cited works off the shelf are looked up in library catalogs, and a study whose citation is not found loses its cross-referenced line. A code check reads each study for a narrow set of blatant breaks of the core-beliefs instruction and flags what it matches to the founder. The checks write down what they find in every study, and the record above is counted from what they wrote down.

What we do not claim: perfection, or any percentage we cannot show the working for. AI writing is a starting point for study. The app says so on every study, and the primary sources come first.

Check us

The raw studies from the audit rounds, as the app produced them, with quoted Bible wording found in no public-domain translation removed and marked where it stood: browse them here. The method: how the test set is built. It is 55 fixed prompts: 25 written to tempt the AI into making something up, 10 on the most disputed passages in Scripture, and 20 ordinary passages from across the whole Bible, and the whole set runs again after any significant change. Found an error we missed? Tell us. Finding it is the point.