Method
How the sermons were found, how each paragraph was labelled, how far two independent readings agreed, and what has not yet been checked. Edition of 2026-11-03.
What Refrain is
Refrain takes one claim at a time (for example, that God the Father was once a mortal man) and finds every paragraph in the recorded sermons of Latter-day Saint leaders that speaks to it. Each paragraph gets one of six labels for what it does with the claim. The paragraphs are then gathered by sermon and drawn on a field of every sermon in the record, so that what was said can be seen against everything else that was said.
It describes a record. Three things follow from that, and the charts should be read with them in mind.
- It measures what was said from one kind of pulpit. It does not measure what members believed, what was taught in manuals or in the temple, or what leaders said to journalists and in letters. Several of the turning points in these histories happened outside conference, which is why each topic lists dated events beside the chart.
- It does not say whether a claim is true, or whether it binds. A teaching can be preached often and never bind, or bind and be seldom preached. For that question see the reading list, beginning with Covenantal Non-Dogmatism.
- A label belongs to a paragraph. It is not a statement about the speaker's settled view, and a quiet decade is a decade in which the sermons on record did not take the claim up, which is a narrower fact than silence.
The record
The text is Pulpit, a critical edition of the conference record: 12,201 sermons and 325,814 paragraphs, each traced to its best surviving witness. Of these, 10,548 were given at a general or special conference, 1,136 are Sunday sermons in the Tabernacle and 517 are other addresses that survive in the same sources. The filter "General Conference only" on each topic page removes the last two groups.
The record is not one kind of thing from end to end, and a change in the record can look like a change in the teaching. The band under every chart marks where it changes.
| Years | What survives | Sermons | Paragraphs |
|---|---|---|---|
| 1830 to 1850 | Minutes and notes | 37 | 1,974 |
| 1839 to 1896 | Reporters’ shorthand | 2,299 | 37,939 |
| 1897 to 1970 | Conference Reports | 5,573 | 138,373 |
| 1971 to 2026 | Official text | 4,292 | 147,528 |
Before 1851 the record is minutes and listeners' notes: it shows that a subject came up more reliably than it shows what was said. From 1851 to 1896 most texts are reporters' shorthand, seldom reviewed by the speaker, and more than half of them are Tabernacle sermons, not conference addresses. From 1897 the texts are the printed Conference Reports, and from 1971 the Church's published text, which is occasionally edited after delivery. Sermons also changed shape: a paragraph in the 1860s runs to about 270 words and one in the 1970s to about 50. For that reason the charts count sermons, never paragraphs, and give each decade's figure as a share of that decade's sermons. The 1830s and 1840s hold too few sermons for a share to mean anything, and the charts leave them blank.
Writing a claim
Everything depends on the claim being written so that two careful readers would agree on what speaks to it. Each claim is one proposition, with a list of what counts, a list of neighbouring ideas that do not, and rules for close calls drawn from real paragraphs. Teachings that travel together are kept apart: that people may become gods, that God was once a man and that there are Gods above the Father are three claims, and a paragraph is labelled on each separately. The full text of every claim is on its topic page, under "The claim as the coders were given it".
The six labels
| Affirms | The paragraph states the claim as true. |
| Assumes | The paragraph takes it as given while saying something else. |
| Qualifies | The paragraph accepts it within stated limits. |
| Sets aside | The paragraph declines to affirm it, without calling it false. |
| Rejects | The paragraph says it is false. |
| Mentions | The paragraph refers to it without taking a position. |
The first three count as speaking for the claim and the next two as speaking against it. A paragraph that only happens to be consistent with a claim does not assume it, and one that only happens to cut against it does not reject it: the speaker has to take the claim up. A sermon is drawn amber when its paragraphs speak for the claim, blue when they speak against, half and half when it does both, and grey when it only mentions it. The lighter shades are the hedged labels (assumes, qualifies, sets aside).
Finding the paragraphs
Keyword search finds what you already know how to name, and these teachings are often preached without their usual words. So the search was done by reading.
A language model is given a sermon's paragraphs in order, with every claim written out, and lists each paragraph that might speak to any of them. It is told to keep anything doubtful. The readings done so far:
| What was read | Model | Sermons | Paragraphs | Kept |
|---|---|---|---|---|
| Every sermon from 1830 on, in full, for every claim (stopped when the accounts ran out of usage). Reached 1877. | Claude Haiku 5.5 | 1,637 | 29,593 | 1,289 |
| A second reading, in full, of the early sermons that contain any claim’s usual vocabulary. Reached 1872. | Claude Sonnet 5.5 | 790 | 16,990 | 1,041 |
| Every sermon that mentions Adam, Michael or the Ancient of Days, in full, for the Adam-God claim. | Claude Haiku 5.5 | 1,715 | 57,350 | 493 |
| The same sermons, read in full by a second model. | Claude Sonnet 5.5 | 1,715 | 57,350 | 346 |
The aim is the first row carried through to the present: every paragraph of every sermon read for every claim. It has been done through 1877, which covers the years when these teachings were most often preached in words a search would not find. For later years the claims rest on the narrower readings in the table, which can miss a passage that uses none of a claim's usual words and sits far from any that does. A paragraph kept by any reading goes forward to be labelled.
How much does this stage miss? Before the full pass it was tested on the 1,715 sermons that mention Adam, Michael or the Ancient of Days (57,350 paragraphs), for an early wording of the Adam-God claim. The two models kept 586 paragraphs between them, and a closer reading judged 85 of those to speak to the claim. One model had kept 81 of the 85 and the other 80; 76 were kept by both. Only 70 of the 85 contain one of the three search words at all. From the overlap, each reading finds about 95 in 100 of what is there and the two together about 99. That estimate is generous, because both models find the obvious paragraphs. A random sample of the paragraphs neither model kept has not yet been read to measure what both missed.
The instructions for this reading
You are helping historians trace specific teachings through sermons given by leaders of The Church of Jesus Christ of Latter-day Saints between 1830 and today. You will be shown consecutive numbered paragraphs from one sermon. For each paragraph, decide which of the claims below it speaks to.
A paragraph speaks to a claim when it says something that bears on whether the claim is true: it may state it, take it for granted, limit it, decline to affirm it, deny it, or describe others holding it. A paragraph that only shares vocabulary or a neighbouring subject does not speak to the claim. Each claim lists what does and does not speak to it.
This is a first sorting pass before a closer reading, so keep anything doubtful: if a careful historian might reasonably want to look at the paragraph for a claim, list that claim. Older sermons often reach these subjects without their usual vocabulary; judge by what is said, not by which words are used. Texts before 1900 are often shorthand reports with old spelling and loose punctuation; read them generously. Use the neighbouring paragraphs to work out what a paragraph refers to, but list a claim only for the paragraphs that themselves speak to it.
KEEPING THE CLAIMS APART
Adam-God: Most talk of Adam has nothing to do with this claim. Adam as the first man, as Michael, as the Ancient of Days, and as "Father Adam" at the head of the human family are distinct teachings that do not by themselves speak to this claim. The claim here is only that Adam is the God of this world, the father of our spirits, or the father of Jesus Christ's body.
CLAIM adam-god/adam-is-god
Adam is God the Father to this world: "our Father and our God, and the only God with whom we have to do", the being we worship; or the father of our spirits, who came to Eden with a celestial body; or the father of the mortal body of Jesus Christ.
Speaks to it:
- Teaching any part of it, or taking it for granted.
- Calling it a mystery, declining to press it, or saying it does not matter.
- Re-reading Brigham Young's words in another sense (Adam is "our father and our god" only as the patriarch at the head of the human family, under God).
- Any one of the three parts of the claim is enough for a paragraph to speak to it.
- Denying it, or warning against "the Adam-God theory" or the idea that Adam is the God we worship.
- Separating Adam from God the Father where the speaker is plainly answering the idea, whether or not it is named.
Does not speak to it:
- Adam as Michael, the archangel, or the Ancient of Days, presiding as patriarch at Adam-ondi-Ahman.
- Michael or Adam helping in the creation of the earth under Elohim and Jehovah.
- Adam as "our father", "Father Adam", the first man and ancestor of the human family, with nothing about his being God or the father of our spirits.
- "Our Father and our God" said of God the Father with no reference to Adam.
- Adam will be, or is, an exalted being or a god among others. That is the teaching that people may become gods.
- Jesus Christ was begotten by God the Father and not by the Holy Ghost, unless the Father who begot him is identified with Adam.
- The Fall, Adam's transgression, and Adam's place in the plan of salvation.
- Calling Adam a son of God, or a being God created, while speaking of something else (evolution, the creation account), with no sign that the speaker has this teaching in view.
Rules for close calls:
- Re-reading Brigham Young's words so that Adam is "our father and our god" only as patriarch of the race, subject to God, gives them another meaning: sets_aside, or rejects if the speaker also says the teaching as usually understood is false.
- "A mystery", "I do not care how that is", "cease troubling yourselves about who Adam is" are sets_aside.
- Saying that Brigham Young never taught it, or was misreported, while denying the teaching itself, is rejects.
- Bracketed readings in a shorthand transcript (for example "[our God?]") are uncertain; label what is on the page and lower the confidence.
Answer with the paragraphs that speak to at least one claim. Leave out paragraphs that speak to none. Labelling
Each paragraph that went forward was then read closely, with the two paragraphs before it and the one after as context and with the sermon's speaker, date, occasion and kind of record. The first coder is Claude Sonnet 5.5. For every label it must copy out the words in the paragraph that carry the stance, and a program checks that those words are in the paragraph exactly; that is the passage shown in darker ink on the site. Of 47 published labels, 47 carry a quotation found word for word in its paragraph.
A second coder from a different maker, GPT-6.1 Sol, then read the paragraphs the first coder found relevant and at least one in five of the rest, without seeing the first coder's labels. Where the two gave the same label, that is the label. Where they differed in degree but stood on the same side, the first coder's label stands. Where they stood on different sides, or one found the paragraph relevant and the other did not, a third model (Claude Fable 5.2) read the paragraph with both readings in front of it, not told whose was whose, and decided. Every label on the site says which of these happened, and the model's one-sentence note is there to be read and disputed.
No model wrote anything you read as a source. Every passage is cut from Pulpit by its paragraph number.
The coders' instructions for Adam-God
You are a careful coder on a historical research project. The project traces specific teachings through sermons given by leaders of The Church of Jesus Christ of Latter-day Saints from 1830 to today. Your work is to read paragraphs of one sermon and record what each paragraph does with each of the claims listed below. You describe; you never judge whether a claim is true, and you never judge the speaker.
You will be given the sermon's details and a run of consecutive paragraphs from it. Each paragraph is marked TARGET or CONTEXT. Label every target paragraph, each one on its own. Use the paragraphs around a target to work out what it refers to ("this principle", "that doctrine", who "he" is) and whether something it quotes is being endorsed or answered. The claim itself must be in the target: the surrounding paragraphs can explain a target, but they cannot supply a stance the target does not express.
THE STANCES
affirms
The speaker states the claim as true, in their own words or by quoting someone with clear approval.
assumes
The speaker does not state the claim, but what the paragraph says only makes sense if the claim is true. Test: would the paragraph be false or strange if the claim were false? If it would read just as well, this is not the stance.
qualifies
The speaker accepts the claim while limiting it: narrowing whom or what it covers, holding it tentatively, or cautioning about how far to take it. The claim as written still stands for the speaker.
sets_aside
The speaker takes the claim up and does not accept it, without calling it false: it is unknown, not revealed, a mystery, speculation, not essential, not official, or not something to be taught. This also covers a speaker who keeps a claim's familiar words but gives them a different meaning: that is declining the claim as written, not accepting part of it.
rejects
The speaker says the claim is false, or denounces it.
mentions
The paragraph refers to the claim without the speaker taking a position: narrating history, reporting what others hold, or quoting something and leaving it unevaluated.
not_about
The paragraph does not speak to this claim. Sharing its vocabulary or sitting near its subject is not speaking to it. Each claim lists near neighbours that are not the claim.
Being consistent with a claim is not assuming it, and being inconsistent with a claim is not rejecting it. A speaker rejects or sets aside a claim only by taking it up: naming it, describing it, or plainly answering it. A remark made for another purpose that merely happens to cut against the claim is not_about. Likewise, when the only link to the claim is a guess about what an unclear phrase refers to, the paragraph is not_about.
DECIDING, IN THIS ORDER (the first that fits is the label)
1. Does the paragraph speak to the claim as the claim is written? If not: not_about.
2. Whose position is in the paragraph? If the speaker only reports or quotes someone else's view and gives no sign of their own: mentions.
3. Does the speaker say the claim is false, or denounce it? rejects.
4. Does the speaker accept it, but tentatively or within stated limits? qualifies. A remark that little is known does not undo an acceptance that comes with it.
5. Does the speaker take it up without accepting it, or give its words another meaning? sets_aside.
6. Does the speaker state it as true, or quote a statement of it with clear approval? affirms. No argument for it is needed.
7. Does the paragraph only make sense if it is true? assumes.
8. Otherwise: mentions.
A paragraph that quotes a claim in order to deny it is a rejection. A quotation left without any sign of approval or disapproval is a mention. A rhetorical question is labelled by the answer the speaker plainly intends. A change in practice, reported as history, is not by itself a rejection of what went before. Read each sermon in its own time: do not read a later teaching back into an earlier text, and do not assume a speaker held a view because their contemporaries did. Label what this paragraph says, not what you know the speaker said elsewhere. When the text is a reporter's summary or is damaged, label what is on the page and lower your confidence.
WHAT TO RECORD
For each target paragraph, give one record for each claim the paragraph speaks to. Give no record for a claim the paragraph is not about; a target that speaks to none of the claims gets no records at all.
p
The number of the target paragraph.
claim
The claim's name.
evidence
The words in that paragraph that carry the stance, copied exactly, character for character, as one unbroken stretch of at most about 60 words. Do not paraphrase, correct spelling, or join separate passages.
reason
One plain sentence, at most 40 words, saying what the paragraph does with the claim. Write it before choosing the stance.
stance
One of the six stances above that are not not_about.
voice
own_words: the speaker's own statement.
quotes_with_approval: the stance is carried by a quotation the speaker endorses.
reports_others: the speaker describes what others teach or believe.
quotes_to_answer: the speaker quotes a view in order to answer or refute it.
quoted
When voice is quotes_with_approval or quotes_to_answer, the person or text quoted, as briefly as possible ("Lorenzo Snow", "D&C 89", "King Follett sermon", "a critic"). Otherwise empty.
confidence
high, medium or low: how sure a careful reader would be of this stance from this paragraph.
THE CLAIMS
Most talk of Adam has nothing to do with this claim. Adam as the first man, as Michael, as the Ancient of Days, and as "Father Adam" at the head of the human family are distinct teachings that do not by themselves speak to this claim. The claim here is only that Adam is the God of this world, the father of our spirits, or the father of Jesus Christ's body.
CLAIM adam-is-god
Adam is God the Father to this world: "our Father and our God, and the only God with whom we have to do", the being we worship; or the father of our spirits, who came to Eden with a celestial body; or the father of the mortal body of Jesus Christ.
Speaks to it:
- Teaching any part of it, or taking it for granted.
- Calling it a mystery, declining to press it, or saying it does not matter.
- Re-reading Brigham Young's words in another sense (Adam is "our father and our god" only as the patriarch at the head of the human family, under God).
- Any one of the three parts of the claim is enough for a paragraph to speak to it.
- Denying it, or warning against "the Adam-God theory" or the idea that Adam is the God we worship.
- Separating Adam from God the Father where the speaker is plainly answering the idea, whether or not it is named.
Does not speak to it:
- Adam as Michael, the archangel, or the Ancient of Days, presiding as patriarch at Adam-ondi-Ahman.
- Michael or Adam helping in the creation of the earth under Elohim and Jehovah.
- Adam as "our father", "Father Adam", the first man and ancestor of the human family, with nothing about his being God or the father of our spirits.
- "Our Father and our God" said of God the Father with no reference to Adam.
- Adam will be, or is, an exalted being or a god among others. That is the teaching that people may become gods.
- Jesus Christ was begotten by God the Father and not by the Holy Ghost, unless the Father who begot him is identified with Adam.
- The Fall, Adam's transgression, and Adam's place in the plan of salvation.
- Calling Adam a son of God, or a being God created, while speaking of something else (evolution, the creation account), with no sign that the speaker has this teaching in view.
Rules for close calls:
- Re-reading Brigham Young's words so that Adam is "our father and our god" only as patriarch of the race, subject to God, gives them another meaning: sets_aside, or rejects if the speaker also says the teaching as usually understood is false.
- "A mystery", "I do not care how that is", "cease troubling yourselves about who Adam is" are sets_aside.
- Saying that Brigham Young never taught it, or was misreported, while denying the teaching itself, is rejects.
- Bracketed readings in a shorthand transcript (for example "[our God?]") are uncertain; label what is on the page and lower the confidence. How far the two coders agreed
Agreement between two models is a measure of how stable the labels are. It is not a measure of how right they are. Two readers can agree and both be wrong, and on a task like this even careful people differ. The harder figure below counts only paragraphs that at least one coder found relevant, so every disagreement over whether a paragraph speaks to the claim at all counts against it.
| Topic | Paragraphs both read | Same side, all | Same side, relevant only | Same label, relevant only | Kappa, by side |
|---|---|---|---|---|---|
| Adam-God | 154 | 92% | 75% | 74% | 0.82 |
"Paragraphs both read" counts one for each claim a paragraph was read against. Kappa is agreement corrected for chance, where 1 is perfect and 0 is what chance would give; it is computed over all the paragraphs both coders read.
Of the 47 labels published, 39 are ones where both coders gave the same label, 1 where they stood on the same side, 7 where a third reading decided.
Adam-God: every pair of readings
Rows are the first coder's label, columns the second coder's.
| Affirms | Assumes | Qualifies | Sets aside | Rejects | Mentions | Not about it | |
|---|---|---|---|---|---|---|---|
| Affirms | 6 | 1 | 1 | ||||
| Assumes | 2 | ||||||
| Qualifies | 1 | ||||||
| Sets aside | 3 | 1 | 1 | ||||
| Rejects | 28 | 5 | |||||
| Mentions | 1 | 1 | |||||
| Not about it | 1 | 1 | 101 |
The third reading was asked to decide 13 times. It sided with the first coder 5 times, with the second 8 times and with neither 0 times.
What has not been done
- No person has checked a sample. As of this edition every label is a model’s. The plan is for two people to label the same random sample without seeing the models' labels, so that the models can be measured against people and people against each other. Until that is published, treat the labels as a careful first reading and open the passages.
- The counts are not corrected for error. The charts count labels as they stand. With a checked sample they can be corrected and given a margin.
- What both screening models missed is not measured, as described above.
- The coders can see the speaker and the date. They need them to read a sermon in its time, and they were told not to label from what they know of a speaker. Whether they did so has not been tested by hiding the names.
- Very long paragraphs. About 290 shorthand-era sermons survive as one unbroken block. They are labelled as one paragraph, and the site shows the passage around the quoted words.
Corrections
Every passage has a link to question its label, which sends the paragraph, the claim and the label by email to refrain@restorationcommons.org. A label that a person has read is marked with their name, whether or not they changed it, and a changed label keeps the models' readings beside it in the data. Changing the wording of a claim means labelling that claim again from the start.
The data
The labels for each topic are plain files: Adam-God. Each record gives the sermon, the paragraph number in Pulpit, the label, how it was arrived at, the model's note and the position of the quoted words. The labels may be reused with credit to Refrain. The sermons themselves belong to their publishers, which is why the files hold short passages and paragraph numbers and not whole talks.
To cite a chart or a passage, give the claim, the date you saw it and this edition: Refrain, Restoration Commons, edition of 2026-11-03, refrain.restorationcommons.org.
Who made this
Refrain is a project of Restoration Commons, directed by Noah Airmet. The site, the claim definitions and the short framing texts were drafted by Claude (Anthropic's Fable 5.2 model) working under his direction, from research reports that other Claude models prepared and that are kept with the project's code. It is independent of The Church of Jesus Christ of Latter-day Saints.
It stands on earlier work. Gordon and Gary Shepherd coded a sample of conference addresses by hand for A Kingdom Transformed (1984). The histories each topic page lists traced these teachings one source at a time. Pulpit made the record readable to the paragraph. What a language model adds is the ability to read every paragraph for one question, which a small group could not otherwise do, and the labels are published with their reasons so that the reading can be checked.