Versioning documents that keep changing: git snapshots, the Darwin variorum, scholarly version rules, and the reasons only an author can record

To keep an honest public history of a document that keeps changing, store every version whole and let software compute the differences. Two things software cannot compute have to be written down when each change happens: why it was made, and where any moved text came from. Git, the version-control system most software is built with, shows the split clearly. It stores each commit as a full snapshot of the project, not as a list of changesKey Findings, and the differences shown to a user are computed only when asked for: The diff a user sees is worked out on demand by comparing two snapshotsKey Findings.

Those who have handled changing texts longest are textual scholars and journal publishers, not programmers. They have been recording reasons and lineage by hand for decades. AI now drafts many of the edits, so the question has moved. A machine can describe what changed. Only the person who accepted the change can say why.

1Git keeps every past state and guesses at relationships between them

The snapshot design suits a public record. Any past state of a document can be shown exactly as it stood, with nothing pieced back together from a chain of edits. The cost is that relationships between pieces of text across time are never stored. Moves, splits and merges between documents are therefore not recorded; they can only be inferred afterwardsKey Findings. That is the report's own inference from how git works, and it follows directly. When a paragraph travels from one paper into another, git holds two snapshots that happen to share some text, and any claim about where the paragraph came from is reconstructed afterwards by heuristic comparison, not written down when the change was made1. What git actually tracks.

For code that rarely matters. For prose it matters more, and prose fits git badly in other ways too. Ink & Switch, a research lab building version control for writers, found that git cannot compare or merge files from WYSIWYG editors like Word1. What git actually tracks, and that Markdown files in git lose features writers rely on, such as threads of comments tied to particular passages1. What git actually tracks. What the lab kept from git was the commit as a deliberate act. In its Upwelling editor, a writer on a private draft chooses how to group related changes into a unit before merging back, even when the changes touch many parts of the document1. What git actually tracks. The unit is defined by the writer's purpose, not by the clock or by the layout of the file.

2Writing has four basic operations, and comparing lines loses the fourth

Literary scholarship has a whole field devoted to how a text came to be. Genetic criticism (critique génétique) treats drafts, deletions and their order as the object of study, not as scaffolding to be thrown away once the book is out. It began in 1968 with the Heine manuscripts acquired by the Bibliothèque nationale de FranceKey Findings. Its home, the CNRS laboratory ITEM, went on to study the manuscripts of Proust, Flaubert, Zola and Sartre.

Its linguists reduced the work of revision to four essential operations in the production of a text: adding, deleting, replacing and moving2. The closer parallel. The linguist Jean-Louis Lebrave pointed out that these are the four basic functions common to all word processing applications2. The closer parallel. Software that compares versions line by line handles three of them. It does not model the fourth, moving, because a move appears as a deletion in one place and an addition in another.2. The closer parallel When an author restructures an argument, moving is usually the operation that carries the most meaning.

The Online Variorum of Darwin's Origin of Species, edited by Barbara Bordalejo, shows both the scale of the problem and a working answer to it. Across the six British editions published between 1859 and 1872, Darwin made around 2,250 sentence-level changes, counting sentences added and removed, and more than 15,000 changes in words or phrases, against a first edition of 3,878 sentences by Peckham's count2. The closer parallel. The specialised collation software the editor used failed in the familiar way: it lost its way when it met long stretches of unmatched text2. The closer parallel.

The variorum's fix was to give every paragraph and sentence a stable number. When the 1869 edition inserted two new paragraphs after paragraph P 100, they became P 110 and P 120.2. The closer parallel A passage keeps its identity across versions even as text is inserted around it.2. The closer parallel The identity is given to the passage when it enters the text. Nobody has to infer it later by matching words, which is exactly where the software broke down.

3Scholarly editions record changes grouped by purpose, and record them on purpose

The digital standard behind such editions, the Text Encoding Initiative (TEI) Guidelines, lets an editor assign individual edits to a change, meaning a state or revision campaign, and give them a sequence number for the order in which they are believed to have happened2. The closer parallel. In software terms, that is a commit grouped by intent rather than by time2. The closer parallel. TEI also openly supports transcriptions whose object is not to represent a finished text, but rather to represent the creative process2. The closer parallel. Its markup can attribute each edit to a hand, with a stated certainty and a responsible party. So an edition can say who made a change as well as what changed.

Word processors made the raw trace richer without anyone planning it. Louis Hay, who led the group that started the field, is quoted saying: Because processors keep a record and the date of every step in the writing process, we can retrace the various drafts and work phases with extreme precision.2. The closer parallel The clearest case is accidental. Researchers recovered Derrida's deleted drafts from his hard disks with forensic techniques, because he did not know the disks kept a trace of his operations2. The closer parallel. A trace like that gives steps and dates. It does not give reasons, and in Derrida's case it was never meant to be read at all. Genetic critics read the gaps between drafts as evidence of thinking, but the thinking has to be reconstructed from outside unless the author wrote it down.

John Bryant's The Fluid Text (2002) gives the theory behind treating the record as content. Against the editorial habit of eliminating variation, he argues that the variations are valuable records of the interaction between artist and society2. The closer parallel. The claim is reported here only through the publisher's description of the book. The report's author did not read the book itself.

4Journal publishers decide which changes count, and require a written account of them

Scholarly publishing deals with what textual scholarship leaves open: identifiers, citation, and how big a change has to be before it counts as a new version. Crossref, the registry behind most scholarly DOIs, keeps drafts out of the citable record altogether: A draft (unpublished) work should not be assigned a DOI3. How scholarly publishing versions a living document. Once a work is published, a new DOI and a public notice are needed only for changes likely to affect how the work is interpreted or credited3. How scholarly publishing versions a living document. For those changes, the original document should not be modified; a separate notice, such as a correction, should be issued instead3. How scholarly publishing versions a living document. That threshold calls for a judgement. Software can count the characters that changed, but it cannot tell whether a change alters what the work claims.

Zenodo, the research repository run by CERN, handles identity with two kinds of identifier. On first publication it registers two DOIs, one for the specific version and one for all versions.3. How scholarly publishing versions a living document It deliberately keeps version numbers out of the identifier itself, on the grounds that this information may change over time, while DOIs must remain persistent and should not change3. How scholarly publishing versions a living document. A quotation can then point to the exact version it was taken from, and a reader who wants the latest thinking follows the identifier for all versions.

F1000Research, a journal where every article can be updated, puts the written account in front of the reader: A short summary of the revisions is displayed at the start of each new version.3. How scholarly publishing versions a living document Its citation format puts the version number and review status in square brackets straight after the title, so they are not lost when the citation is copied3. How scholarly publishing versions a living document.

Its rules for living systematic reviews, defined after Cochrane as a systematic review that is continually updated, incorporating relevant new evidence as it becomes available3. How scholarly publishing versions a living document, are the most fully worked-out rules for a document meant to keep changing. Each new version must include a summary of the differences from the previous one, at most 300 words, published in an amendments box3. How scholarly publishing versions a living document. A second box gives the dates of the last and next literature searches and what the last one found. A review can also lose its living status. Grounds include missed updates, drift from the original question, or reaching the end of its planned schedule3. How scholarly publishing versions a living document. Credit is handled across versions too. Authors who drop off over time are kept in the author lists of earlier versions and credited in acknowledgements, so credit is preserved across versions3. How scholarly publishing versions a living document.

Each of these rules depends on a sentence a person writes: the amendments box, the correction notice, the statement that a search found nothing new. Publishers built them because a stored version, however complete, does not explain itself.

5Readers still prefer to cite a fixed version, so a living document has to be labelled

Lisa Janicke Hinchliffe's 2022 review describes scholarly publishing moving toward a linked record of versions. It also explains why the single fixed version keeps its hold: As the version of record comes into being, it has the effect of fixing the earlier versions as predecessors to it.3. How scholarly publishing versions a living document The preference shows up in the survey she cites.

The two labels measure different things. A document can be finished but low-confidence, or a draft built on solid evidence.4. How writers on the web show that a piece is unfinished A history of versions shows when either one moved.

6When an AI drafts the edits, the written reason is the human's share

Ink & Switch's next project, Patchwork, started from the observation that collaborators now include AI tools proposing edits1. What git actually tracks. The lab also tested the obvious shortcut of letting a language model write the change notes. It found that auto-generated LLM summaries of edits served text well but were "less obviously good" for visual edits1. What git actually tracks. A model can read two snapshots and say what is different between them. What it cannot read off the snapshots is the decision: why one draft was kept, why a claim was softened, why two papers became one.

Research on AI-assisted software work has reached the same split from another direction. Conversation with a model is a poor place to keep decisions. One paper argues that chat is not a stable external representation. It is a transient streamPart 2 — Stigmergy and artifact-mediated direction of AI agents. The model forgets between sessions while the person does not. A 2026 preprint, which is single-authored and unreviewed, describes each session opening with the participant restoring, from artefacts he maintains, a common ground the other party cannot carryKey Findings. Margaret-Anne Storey gives the missing piece a name: Intent debt lives in artifactsPart 2 — Stigmergy and artifact-mediated direction of AI agents. She defines it as the absence of externalized goals, constraints, and rationale that both humans and AI systems needPart 2 — Stigmergy and artifact-mediated direction of AI agents. A version history that stores only snapshots builds up exactly that debt. It has every state of the text and none of the reasons.

There is also a reason to want the dates. AI-drafted work tends to look better at first than it turns out to be. In a Stanford study, researchers spent over 100 hours each carrying out ideas, and after that for many metrics there is a flip in rankings where human ideas score higher than LLM ideasKey Findings. The person judging is not neutral either. In a preregistered experiment, an "Aha!" moment triggered by solving an unrelated anagram made false statements feel truerKey Findings. A dated record of each revision, with a line saying why it was made, is where both effects would show: the AI-drafted claim that shrank on a second reading, and the author's change of mind once the first excitement wore off.

7Merges and splits are where every published scheme stops

The hardest case is structural: two documents combined into one, or one split into several. Crossref and Zenodo link versions of a single work, and Crossref links distinct works such as a preprint and its published version, and the report infers from what both describe that the link a merge needs is missing. Neither scheme is designed to say that one new document supersedes two others5. Splits, merges and machine-readable history. A search for one found no standard identifier scheme recording that one living document superseded two others, or split into severalOpen questions. That gap comes with a stated limit: the full relation-type vocabularies of DataCite and Crossref were not read in full.

The nearest models again come from manuscripts and from the wiki. Genetic editing sets out to reconstruct exactly which parts of the manuscript were acted upon where multiple manuscripts have been combined5. Splits, merges and machine-readable history. In Ward Cunningham's Federated Wiki, every page carries its own edit history, called the journal4. How writers on the web show that a piece is unfinished, and the journal travels with the page when it is copied. Both point the same way as the Darwin numbering. Lineage survives when it is attached to the text at the moment the text moves. Whether stable passage numbers hold up under deep restructuring has not been tested. The Darwin scheme covered editions of one book, and identities may need to be assigned by hand at the moment of a merge.

8Written notes about the work do not automatically help

The strongest evidence against relying on written notes comes from software. When ETH Zürich researchers tested the prose files developers leave to guide AI coding agents, context files produced no improvement in task success while increasing inference cost by over 20%Key Findings. The agents did follow the files, but unnecessary requirements from context files make tasks harderPart 2 — Stigmergy and artifact-mediated direction of AI agents. The research's own reading, which it labels as its synthesis and not a finding, is that traces help only when the traces are constraining rather than descriptivePart 2 — Stigmergy and artifact-mediated direction of AI agents.

The finding concerns notes written as instructions for a machine doing the next task. A change note in a document's history is written for a reader looking back, which is a different job. Still, the warning applies: more prose about the work is not better in itself. The publishers who require reasons also limit them, as with the 300-word amendments box and the single dated line on the last search. A reason earns its place when it says something that cannot be recovered from the text.

The rest of the evidence is thinner than the design ideas built on it. The definition of genetic editing used here comes from an encyclopedia overview, because the primary book was not available. Bryant's argument is known only through his publisher's description. On readers, Hinchliffe notes that systematic research on how researchers view versions is scarceOpen questions, and the studies she reviews concern preprints, not documents that are revised continuously. She also expects the old arrangement to hold. Having sketched a future without any single fixed version, she ends by concluding that the fixed version of record will probably remain central3. How scholarly publishing versions a living document. That does not argue against keeping a history. It argues for keeping fixed and living documents clearly apart.

9The reason for each change is where authorship shows once machines draft the text

Taken together, the traditions agree on a division of labour. Storing versions and comparing them is a solved mechanical problem, and git solves it well. Every tradition that treats the history of a text as worth reading adds something the machinery cannot supply: TEI's revision campaigns and hands, the variorum's passage numbers, Crossref's notices, F1000's amendments boxes and preserved author lists. Each is a fact about intent or lineage, written down when the change was made because it cannot be recovered afterwards.

With AI drafting much of the text, that division becomes a question of authorship. The snapshots show what the machine proposed and what the document became. The dated line saying why a change was kept, and by whom, is the only place the human judgement in between leaves a mark. A changing document that records it shows readers where the author's thinking happened, not only how the text ended up.