# Different writers, shared traditions *How stories, knowledge and error travel through memory, music and machines* *By Luuk Joseph Soons · 26 October 2026* How can a tradition change without losing its memory? Christopher Nolan's 2026 adaptation of *The Odyssey* gives an ancient story another telling. A cinema audience encounters a new film; a reader may reach for an old book. Behind both lies a history of poetry carried by human voices. The journey from song to screen invites a question that now reaches beyond literature: what survives when the people, forms and instruments of transmission change? [Universal Pictures, *The Odyssey*](https://www.universalpicturesathome.com/movies/the-odyssey-2026). The Homeric poems grew from oral poetic traditions, but scholars disagree about how they acquired their written form: Gregory Nagy describes a gradual process of performance and textualisation, Richard Janko argues for oral performances taken down by dictation. We cannot open an author's surviving manuscript and measure every later departure from it; nor does uncertainty about that beginning establish that there was never a particular act of composition. [Nagy, “Performance and Text in Ancient Greece”](https://www-current.chs.harvard.edu/curated-article/gregory-nagy-performance-and-text-in-ancient-greece/); [Janko, “The Homeric Poems as Oral Dictated Texts”](https://discovery.ucl.ac.uk/id/eprint/173855/1/download3.pdf). Writing did not make every subsequent copy identical. Ancient *Odyssey* papyri at the University of Michigan preserve differences in wording, and one fragment lacks a verse found in the standard text: continued variation, though not a measure of how much the whole poem changed. Such differences within the Greek tradition are also distinct from a translation or a film, whose maker deliberately creates another work from the received poem. [University of Michigan, *Translating Homer: From Papyri to Alexander Pope*](https://apps.lib.umich.edu/online-exhibits/exhibits/show/translating-homer--from-papyri/translating). A stable text can support a living tradition. It gives later readers something to compare, question and reinterpret. Without a preserved record, an alteration may pass unnoticed; without renewed interpretation, preservation can become mere repetition. An institution devoted to knowledge needs both a dependable record and the freedom to reconsider what it means. The question has another setting today. A person writes a passage. A language model rewrites it. An editor corrects the result, which is then passed to another model. The participants differ, but their texts belong to a connected history. Calling one version human-written and another AI-generated tells us something about their production. It leaves much of their relationship unexplained. What did each participant receive, preserve and make anew? In the South Slavic epic traditions studied by Milman Parry and Albert Lord, singers could compose during performance using inherited formulas and themes. They learned resources for making a song, rather than simply memorising an unchanging script. A written record captured one performance; it did not necessarily become an authoritative original for the continuing tradition. Lord's account therefore gives us a reason to distinguish preservation of a tradition from repetition of identical wording. These observations concern particular epic practices, not every oral culture or genre. Oral traditions preceded writing and continued alongside it. [Lord, *The Singer of Tales*, chapters 1 and 6](https://chs.harvard.edu/book/lord-albert-bates-the-singer-of-tales/). Nor does working within a tradition extinguish individual invention. Lord's later discussion gives a place to exceptional performers and creative elaboration as well as recurring formulas. The familiar and the newly made can belong to the same performance. [Lord, “Characteristics of Orality” (1987)](https://journal.oraltradition.org/wp-content/uploads/files/articles/2i/6_lord.pdf). Music helps explain how words can endure. A rhythm can cue the next phrase; rhyme can narrow the words that fit. David Rubin's work on oral traditions examines how patterns of meaning, imagery and sound constrain recall, supporting memory while shaping what is remembered. The benefit is conditional: in Wanda Wallace's experiments, people recalled words better when several verses shared a melody, but with a single sung verse, or different melodies across verses, spoken presentation worked better. Music helped when its structure could be learned and reused. [Rubin, *Memory in Oral Traditions* (1995)](https://academic.oup.com/book/53811); [Wallace (1994)](https://doi.org/10.1037/0278-7393.20.6.1471). For an investigator of transmission, this raises a difficulty. Familiar forms can help a passage survive, yet they can also make separate performances resemble one another. Two singers may draw on the same repertoire without either having heard the other's song. A recurring expression may identify a tradition more readily than a particular predecessor. We must decide what history we are trying to recover: the history of a phrase, a narrative episode, a performance, or the resources available to its maker. This is where oral tradition usefully complicates stemmatics, the study of relationships among surviving versions of a work. If we assume a fixed original too quickly, we may classify creative variation as damage. If we treat every difference as invention, we may overlook what has been inherited. Comparison needs an account of the activity that produced the versions, including what participants were expected to preserve. The same care is needed when we turn to language models. A singer and a model need not work alike for their outputs to raise comparable questions about transmission. The comparison concerns what can be learned from successive versions. It does not establish a common cognitive mechanism, or make intention and responsibility interchangeable. A recent study by Yngwie Nielsen and Morten Christiansen prompted this inquiry. Their experiments found evidence that people process recurring grammatical sequences that do not correspond to conventional syntactic constituents, while leaving room for constituent structure in larger sequences and full sentences. That is a narrower reason to attend to familiar sequences in language processing, not a finding that human language has no hierarchy or that brains work like language models. [Nielsen and Christiansen (2026)](https://www.nature.com/articles/s41562-025-02387-z). Transmission research gives the inquiry firmer experimental ground. In a study of successive story retellings, Fritz Breithaupt and colleagues found that human versions changed incrementally, while ChatGPT tended to condense material early and make fewer subsequent changes. The comparison has a limit visible in the study's method rather than stated by its authors: people were asked to remember the story and retell it, whereas the model, prompted to retell, presumably had the text before it, a condition the paper does not describe. A transmission task can be studied across different participants without their capacities or conditions being equivalent. [Breithaupt et al. (2024)](https://pmc.ncbi.nlm.nih.gov/articles/PMC10776760/). Other experiments show why resemblance needs careful interpretation. Jérémy Perez and colleagues found that repeated transformations by language models could move texts towards characteristic patterns in positivity, toxicity and length, depending on the model and the instructions. Similar outcomes may therefore reflect recurring pressures on the text as well as its starting point; the result concerns measured properties, not the ancestry of any particular passage. [Perez et al. (2025)](https://proceedings.iclr.cc/paper_files/paper/2025/hash/dbdea7859f1d2fc10f2c9e79b8f5ae54-Abstract-Conference.html). Human authorship offers no automatic escape from shared influence. In Anil Doshi and Oliver Hauser's short-story experiment, writers offered AI-generated ideas produced stories that evaluators rated more highly on average, with the gain concentrated among the less creative writers, and the stories were collectively more similar. A bounded writing task, not a verdict on creativity; but it shows why separate human names cannot, by themselves, establish separate sources of inspiration. [Doshi and Hauser (2024)](https://doi.org/10.1126/sciadv.adn5290). Consider a hypothetical case. A researcher mistakenly gives an agreement's date as 1847 rather than 1849. A model summarises the passage and retains 1847. An editor improves the prose without checking the date. A later model quotes the edited version. Four documents now carry the same mistake, but they need not represent four independent judgements. Their agreement may be the continuing life of one error. Even this simple example requires evidence. A record showing that a passage appeared in a prompt establishes access. It does not, on its own, establish which details were taken from it. To argue that the wrong date travelled along the proposed route, we would want the intermediate versions, the material available at each step and a comparison with other plausible explanations. Another source might contain the same date. The error might also be a common one. A distinctive mistake can be useful evidence, but rarity alone is insufficient. The mistake should be more likely to appear through the proposed transmission than through the alternatives we are considering. A peculiarity that almost never arises independently, but is even less likely to survive copying or rewriting, would not support that transmission merely by being unusual. What matters is the comparison between explanations. Nor must every informative feature be a mistake. A peculiar ordering of examples, an unusual metaphor or an added qualification may help us trace a passage. Whether a change is an error depends on the task. Changing the date in a faithful historical summary is different from changing a place name while adapting a story. Before counting departures, we need to know what fidelity required. These distinctions have consequences for AI-assisted review. Several assessments may share a conclusion because each found adequate evidence. They may also share a source, an instruction or an earlier assessment that none successfully challenged. The number of voices does not settle how much additional evidence their agreement supplies. Conversely, showing that two texts share a history does not establish that their conclusion is false. Provenance bears on how we weigh testimony; a claim's truth requires its own grounds. Historical source criticism has long confronted this problem. In their 1898 *Introduction to the Study of History*, Charles-Victor Langlois and Charles Seignobos distinguish the number of documents from the number of observations behind them. A hundred reports copied from one account do not supply a hundred observations. Their discussion also asks us to examine assertions separately, rather than accept or reject a document as a whole. These are useful precedents for assessing AI testimony, although assumptions about a human witness's sincerity cannot simply be transferred to a model. [Langlois and Seignobos, Book II, especially chapters VI to VIII](https://www.gutenberg.org/cache/epub/29637/pg29637-images.html). This places stemmatics beside a broader inquiry. Source criticism asks what grounds a statement has, given how its source was produced. Hermeneutics attends to meaning in context: what did “checked” or “confirmed” mean in this particular account? Stemmatics asks how the surviving texts are related. Their questions overlap, but answering one does not settle the others. There is an institutional question here too. I imagine CounterProof Research as a scientific monastery. Its monks are temporary agents: a session begins, takes up a task and eventually ends. Before leaving, it can preserve its sources, work and unresolved questions for a successor. The institution maintains the library, rules, methods and tools through which these brief episodes become continuing work. The metaphor concerns that continuity. The session is temporary; the scientific practice persists. The parts have ordinary functions beneath their technical names: a library preserves material; retrieval selects passages to place before a model; skills supply reusable instructions; a harness arranges tools and the sequence of work; a handover says what was attempted and what remains uncertain; human stewards set the purposes and remain accountable. Research on retrieval and reports on agent harnesses support benefits in particular settings, not a general rule that more context or machinery improves judgement. [Lewis et al. (2020)](https://arxiv.org/abs/2005.11401); [Anthropic, “Harness design for long-running application development” (2026)](https://www.anthropic.com/research/harness-design-long-running-apps). Two established literatures bear on this arrangement. Edwin Hutchins studied navigation as cognition distributed across people, instruments and practices, with properties no single navigator explains; applied to agents, that perspective means studying the whole arrangement through which a task is accomplished, human supervision included, without attributing human experience to the machines. [Hutchins, *Cognition in the Wild* (1995)](https://mitpress.mit.edu/9780262581462/cognition-in-the-wild/). James Walsh and Gerardo Ungson examine how organisations acquire, retain and retrieve information, and how that memory is used or misused: a neighbouring literature for asking what survives changes in personnel. In an agentic institution much of this inheritance is supplied afresh through documents, retrieval and tools; a new session can use the same model as its predecessor in a better environment, and the model's weights need not change for the institution's conduct to improve. [Walsh and Ungson, “Organizational Memory” (1991)](https://doi.org/10.5465/AMR.1991.4278992). Preservation and learning nevertheless describe different achievements. Suppose an agent relies on an outdated source. Its transcript records the mistake. A correction identifies the better source and explains the difference. A revised retrieval rule or skill aims to help later agents avoid the same mistake. If successors using that correction handle fresh cases better than successors working without it under comparable conditions, we have evidence of useful institutional learning. A larger archive alone would not show it. Nor would success on the exact example from which the new instruction was written. Successors should not have to rediscover every established result. The point of inheritance is to let work accumulate. They need enough of the evidence, reasoning and scope to judge when a result remains usable and when changed circumstances require reopening it. A correction should travel with the claim it replaces. A handover should preserve unfinished questions alongside achievements. In [*Earned, Not Inherited*](https://research.counterproof.io/essay/earned-not-inherited.html), I express the principle as “share the standards; earn the verdict”. What is inherited gives the next investigator somewhere to begin. This makes distinctive AI-assisted institutions plausible. Two engineering firms could use the same model yet develop different practices because they preserve different cases, demand different evidence and learn from different failures. A creative studio might acquire a recognisable taste through its selection and revision of work. Such traditions would reside in the continuing relationship among people, records and tools. Their persistence across changes of model, and any improvement they bring, remain questions to investigate. A recognisable house style would establish something different from greater accuracy. A common culture can also accommodate several model providers. Models from several providers can share standards for evidence while examining a case separately. When a fresh assessment is required, the institution should supply the relevant methods and source material while withholding the previous verdict until the first reading is recorded. That boundary is task-dependent: an agent continuing unfinished work may need the very handover a fresh reviewer should not see. Shared methods, useful inheritance and independent assessment need different arrangements, rather than one rule about sharing everything or nothing. Agentic stemmatics can investigate what moves through those arrangements. A record may connect a session to a handover, the handover to a memory entry, and that entry to a successor's context. It establishes a route of access; whether a particular detail was taken from that source remains a further question. Where records are missing, textual comparison may help infer relationships, provided the method is tested against known histories and plausible alternatives. The handover raises another question: what account does it make of the work? Frank Ankersmit's philosophy of historical representation distinguishes the statements in a text from the representation they form together, and directs attention to the aspect of a subject that an account presents. Applied here, it asks how selection, arrangement and emphasis turn a collection of records into an institutional account; an application we propose, not a claim that his work establishes a theory of AI reports. [Ankersmit, “Six Theses on Narrativist Philosophy of History”, in *History and Tropology* (1994)](https://publishing.cdlib.org/ucpressebooks/public/book/history-and-tropology-the-rise-and-fall-of-metaphor.html); [Ankersmit, “Representation and Reference” (2010)](https://research.rug.nl/en/publications/representation-and-reference/). Imagine that three agents agree and a fourth identifies a consequential mistake. A summary could foreground agreement or the correction. Even with accurate quotations, it could leave a successor with a misleading impression of what the review accomplished. The proper emphasis depends on the task and the evidence. A short account cannot include everything; the question is whether its omissions change what a reader is reasonably entitled to conclude. Ankersmit should not be invoked to excuse factual error. His 2012 discussion expressly allows historical truth while distinguishing the truth of statements from that of a representation. His later *Representation* extends his account of historical argument. We need to engage that work critically, including objections such as John Zammito's argument that historical organising concepts remain answerable to claims about reality. For our purposes, a useful assessment would examine both the supported statements and the overall conclusion they are made to carry. [Ankersmit, “Truth” (2012)](https://academic.oup.com/cornell-scholarship-online/book/20388/chapter-abstract/179512713); [*Representation* (2024), publisher's account](https://cup.columbia.edu/book/representation/9780231561037/); [Zammito (2005), on Ankersmit's earlier position](https://doi.org/10.1111/j.1468-2303.2005.00316.x). The promise of new discovery deserves the same distinction. Models can propose connections across disciplines, but an illuminating resemblance begins as a suggestion and becomes a contribution only when it survives the relevant mathematics, measurements or other checks. FunSearch paired a language model with a systematic evaluator and found new mathematical constructions and useful algorithms; its success depended on problems whose candidate solutions could be checked. That supports a possibility, not a licence to treat fluent synthesis as discovery. [Romera-Paredes et al., “Mathematical discoveries from program search with large language models” (2024)](https://www.nature.com/articles/s41586-023-06924-6). New models may help us sing new songs, but science asks more of a song than coherence. An invented episode can belong in an adaptation of the *Odyssey*. An invented experimental result cannot belong in a faithful account of research. A research institution needs room for conjecture and a clear account of when, why and within what limits a conjecture earned belief. Otherwise its accumulated literature may become an increasingly elegant account of mistakes it has never tested. In [*Counterpoint in music, CounterProof in method*](https://research.counterproof.io/essay/counterpoint-and-counterproof.html), I used musical voices to explore the value and limits of distinct contributions. Counterpoint directs attention to voices heard together; transmission directs attention to versions carried through time. The two are complementary, provided we remember that melodic independence does not demonstrate independence of origins, judgement or error. The companion note [*Collation and the microscope*](https://research.counterproof.io/note/collation-and-the-microscope.html) asks whether comparisons of model outputs and measurements of their internal workings could check one another; the tests it sketches, and the eight conditions it says no design yet meets, concern specified relationships between models. Interpretability can investigate mechanisms that comparisons of documents cannot reveal, though its instruments have limits of their own: circuit-tracing work on Claude 3.5 Haiku describes only parts of the computation it studies, and a model's own explanation is no substitute for that investigation. The present essay concerns relationships between passages, performances and institutional practices. Passing a passage through a model does not ordinarily change its weights, so a text's transmission history must not be mistaken for the model's training history; nor would agreement between an internal measurement and a textual comparison validate either by itself. [Lindsey et al., “On the Biology of a Large Language Model” (2025)](https://transformer-circuits.pub/2025/attribution-graphs/biology.html). These questions suggest three neighbouring research programmes within CounterProof Research. Distributed agentic cognition and organisational memory would study how work accumulates across sessions, people and tools. Agentic representation would study how records become summaries, explanations and institutional accounts. Agentic stemmatics would investigate transmission: the evidence for relationships among texts and among the models that produce them, and how far the errors of reviewers depend on one another. They are programmes drawing on established fields, with separate questions and standards of success, not three new disciplines; their shared home is the institution, and each would be judged on its own results. The institutional question could be tested by comparing the same model on comparable tasks under different memory arrangements, distinguishing ordinary retrieval from a curated memory that carries corrections and their reasons; fresh cases, comparable resources and a record of mistakes would show whether a procedure teaches something reusable. Whether another model adds enough useful work to justify its cost is a different question. The representation question would ask whether examining the whole account adds value beyond checking its individual claims and citations: summaries compared against preserved case records, including adverse or unresolved evidence, with misleading omissions distinguished from reasonable compression and unjustified objections counted alongside real catches. An account does not become adequate because several models approve it; the relation between its evidence and its conclusion stays open to human examination. The transmission question would ask whether a reader or model can identify which passage was supplied for the next rewrite, against what simple textual similarity already recovers. A successful institution need not reconstruct textual histories well, and a successful reconstruction method need not produce a better summary; keeping these outcomes separate lets each programme find its own limits. A first transmission test could use models alone: short passages passed through different sequences of models, every input and output retained, each reconstruction attempted in a fresh session from anonymous texts. The difficult cases would matter most. Two rewrites can share a source without either being copied from the other; two compositions can arise from one brief. The test must allow the answers that no candidate was supplied or that the evidence is inconclusive, and count false links and abstentions alongside correct answers; a method that sees ancestry everywhere may draw an impressive genealogy while getting the history wrong. Later stages could add people, with retelling from memory and rewriting from an open source kept apart, and a separate study could ask what rhythm and melody do to recall. Neither transmission chains nor known-history tests are new: handoffs between people and language models appear in Acharjee and colleagues’ study of information fidelity, and artificial textual traditions have been used to evaluate stemmatic reconstruction. A study along these lines would ask, under specified conditions, how well inherited features can be distinguished from independently recurring ones; its baselines and staged conditions belong in a registration rather than here. It could succeed in one setting and fail in another, and if a richer method added nothing to a simple comparison of wording, that too would be a result. [Acharjee et al. (2025)](https://aclanthology.org/2025.ijcnlp-long.146/); [Roos and Heikkilä (2009)](https://doi.org/10.1093/llc/fqp002). There is a further limit that no reconstruction can remove. A story can be faithfully transmitted and remain false. A musical pattern can help preserve an inaccurate claim too. Establishing who inherited a claim is therefore different from establishing whether it should be believed. In code review, an executable test may sometimes settle a particular allegation. Where such a check is unavailable, the question of how much confidence the testimony deserves remains harder. The *Odyssey* offers a fitting return. Odysseus is admitted home not on his own account of himself but on marks the household can test: a scar the old nurse knows, a bed that cannot have been moved because it is built on a living olive tree. The returning voice is heard, and then it is checked. A preserved text and a new telling can enrich one another without becoming interchangeable. Institutions that work with AI will need a comparable ability to retain what was said while allowing better accounts to emerge. Their character will be formed by what they keep, what they question and what they are willing to withdraw. The scientific monastery gives that ambition an institutional form. Its agents arrive and depart; their work can remain available to correction and use. Stemmatics may help recover its transmission, representation may help examine the accounts built from it, and research on organisational memory may help turn experience into better practice. Each contribution must earn its place. The institution's purpose is to leave its successors better equipped, while preserving enough of its past for them to understand both its achievements and its mistakes. *This is an exploratory essay. The transmission and handover examples are hypothetical. The institutional, representation and transmission studies remain proposals; no completed experimental findings from them are claimed. “Tradition” and “scientific monastery” describe continuity and revision in shared practices, not a claim that models possess human consciousness or cultural experience. The cited studies support particular claims within their stated settings; they do not establish cognitive equivalence between humans and machines or validate the reliability of a commercial review service. The Ankersmit discussion draws on the original text of the 1994 “Six Theses” and on published abstracts of the later work; the 2024 book is introduced through its publisher's account, rather than represented as a completed reading. The author has a commercial interest in CounterProof Research, a practice that sells adversarial review services; the research programmes proposed here, if they succeed, would serve that business. This essay was drafted and revised with Codex (OpenAI); the revision of 26 October incorporates the author's clarification of the scientific monastery and the proposed separation of the research programmes.* --- *CounterProof Research essay, 26 October 2026. By Luuk Joseph Soons, director and co-founder of CounterProof Research, a practice of Clavestra Capital Limited. Drafted and revised with Codex (OpenAI) and reviewed before publication by Grok (xAI), under an adversary brief, and Gemini (Google), each under its own brief; their rows were adjudicated by GLM (Z.ai) and the applied changes confirmed by Grok in two rounds. The review records are held in our reports store and available on request.* *[CounterProof](https://counterproof.io/) is a small practice of Clavestra Capital Limited (Malta, C 113987) that reviews AI-generated code and AI-assisted review processes for organisations where a wrong answer is expensive. We disclose that we have a commercial interest in the adoption of the review methods this essay discusses.* *Cite as: Soons, L. J. (2026). Different writers, shared traditions. CounterProof Research, 26 October 2026. Companion to [Agentic Stemmatics](https://research.counterproof.io/agentic-stemmatics.html), doi:10.5281/zenodo.22030516.*