The Recursive Book

The first preserved words of this book are an instruction. They ask for two things at once: a readable book about machine-assisted literary production, and a way of making the book that leaves important parts of its production open to inspection. The instruction also supplies a proposition to test. It proposes that a maximally self-referential machine-written book does more than describe its production: it alters later processes involved in its meaning, authorship, interpretation, and regeneration.¹

On the page, the instruction resembles a first cause. Before there is a chapter, there is a command to make one; before there is an argument, there is a sentence naming what the argument must investigate. But a proposition can govern an inquiry without being its result. The commission establishes the question and its burden. It does not establish that the proposed effect occurs.

The prompt is therefore a beginning of a particular kind. It is the earliest fully preserved commission in this project’s governed record. It is not the first cause of its language, its concepts, its technical setting, or the events that made the commission possible. The visible beginning is already late.

The surviving record attributes the commission, its stated problem, and its constraints to the supervisor. It does not preserve the first conception of that problem or establish which earlier conversations or decisions preceded the supplied wording. What it does show is an address: a human supervisor supplied a task to a recorded system activity, and an answer followed.

The words of that address arrived later than the language in which they were written. They carry genres, concepts, and phrases with histories that neither participant began. Some prior texts can be named because the project quotes them, cites them, or uses them to frame a question. Its broader linguistic inheritance remains unenumerated. Dependence on prior language defeats the fantasy of expression without antecedents; it does not turn every earlier sentence into an author or reveal a source-specific training path.

Technical and institutional layers also precede the answer, but the evidence is uneven. The first generation record reports a Codex desktop session and a GPT-5-family model identity while leaving the exact served version, full runtime context, and the decisions behind model construction, deployment, access, and interface design outside the record. Retained studies of other named systems distinguish dataset description, corpus construction, and relations to training runs. No comparable chain here identifies the items consumed by this served model or the influence of a particular source on a particular sentence.²

These are not invitations to invent the missing history. They are differences in what can be said. A named source relation is not an acting contributor. A reported technical setting is not its complete institutional biography. An unknown lineage is not a secret chamber inside the machine. It is an unknown relation.

What survives is partial but not empty: an instruction, an answer, rejected paragraphs, changed sentences, named sources, objections, and decisions to retain or revise. Such traces can distinguish some events and participants. They cannot reveal hidden token selection, recover every element of context, or turn chronological order into a complete causal chain.³

The prompt begins the inspectable inquiry without beginning all of its conditions. To say that it has causes is not to give every antecedent the same verb. Prior texts do not decide. Interfaces do not commission. Conditions do not become actors merely by being present. Recorded actors do not acquire credit or responsibility merely because their actions came earlier.

The plural therefore creates a problem rather than a crowd byline. Which antecedents are relevant background? Which are operating conditions, source dependencies, attributed actions, selections, or bounded contributions? Under what additional rules, if any, should one of those relations count toward credit, responsibility, or authorship? Before asking who wrote what follows, we have to ask what kind of cause this prompt was - and what kinds it was not.

Return to contents

The request and the answer

The preserved instruction asks for a book. It supplies a title, a subject, a candidate proposition, a method, and an unusual permission: the writing system may continue with considerable autonomy while preserving enough of its work for later inspection. A recorded activity then emits candidate prose. Later, other activities reject, select, and revise what is available. The reader encounters what those later decisions retain.

This short chain already contains unlike relations. The commission states a task. The technical setting permits an exchange. A generation supplies wording. Retention determines which available wording becomes part of the manuscript. Calling all four events causes would be harmless only if the word kept their differences visible. Calling all four authors would settle a question the chronology has merely raised.

The differences alter what each event can explain. The instruction can explain why the project investigates self-reference and why it must preserve a readable manuscript beside an evidence practice. It cannot explain why a later draft chose one image, cadence, or transition rather than another. Emission can explain where candidate wording first appears in the retained artifacts. It cannot explain why that wording survives criticism. Retention can explain why a reader meets the survivor. It cannot claim the survivor’s entire prehistory as its own. A single origin story loses these distinctions exactly where the work becomes interesting.

The instruction has real force. It gives the project an identity and a governing question, requires both readable prose and an inspectable production practice, and authorizes research and revision. It does not contain every sentence that will answer it. It does not establish which candidate will survive or whether its proposed thesis is true. The commission is evidence of direction at a recorded scope, not a complete blueprint and not independent proof of the proposition it commissioned.¹

An observable instruction entered an already configured situation and precedes the governed project sequence preserved here. That is a precise statement of order. It does not yet tell us how much the prompt explains, which of its antecedents affected a particular sentence, or why any participant should receive credit. To answer those questions, the history needs more than a backward list.

Causes answer questions

The word cause invites regress. Move backward from the instruction to the person who supplied it, from that person to prior texts and institutions, from the technical event to the conditions behind it. Continue far enough and the history of a sentence becomes the history of the world. Nothing has been omitted, in principle; nothing has been explained, in practice.

A cause becomes relevant here as an answer to a bounded question. Ask why this project has this title and subject, and the commission is central. Ask why an answer could be recorded, and the reported operating setting comes forward. Ask why a concept appears in an argument, and an identified source or research instruction may matter. Ask why one available paragraph rather than another reaches the reader, and a later retention decision matters. Each question selects a different part of the history. An answer to one does not automatically answer the others.

Four provisional descriptions help without pretending to be a universal taxonomy. An available interaction system is an enabling condition of the recorded exchange. Requirements about form, sources, roles, and prohibited claims are designed constraints. The commission’s expressed subject and objective are task directions. Languages, genres, locutions, and arguments already available to the project are inherited linguistic or textual resources. In the actual work, these descriptions can overlap. More importantly, none names a kind of author.

The labels restrain two temptations. First, chronology is not attribution. A condition may precede a passage without the record isolating any effect it had on that passage. Second, relevance is not necessity. The repository contains no registered rerun of the complete production history with each antecedent removed in turn. An imagined alternative can identify a question for testing; it is not an observed result. The book can say that a rule was present, a source was cited, a task was supplied, an answer was emitted, or a paragraph was retained. Stronger causal verbs require stronger evidence tied to a specified target.²

The backward search stops at the boundary of the question, not at a claim that earlier causes cease to matter. The chapter selects antecedents that clarify the differences among commission, setting, textual resources, generated alternatives, and the artifact a reader receives. A partial explanation that states its selection is more useful than a total inventory that cannot say why any item belongs.

Two solitary origins

Two convenient stories abolish these distinctions by giving the work a sovereign origin.

The first places sovereignty in the machine. Prose appears with a coherence and velocity that make its dependencies easy to ignore. Because a recorded system activity emitted the visible wording, emission is expanded into a whole authorship story. But the output did not supply its own task, construct the setting in which the exchange occurred, or establish the standards by which its claims would later be retained. The record supports an attributed act of generation. It does not support a self-originating machine author.

The rival story places sovereignty in the human commissioner. The model becomes a transparent instrument; the person who issued the prompt is treated as having written the responses through it. Yet supplying a subject, an objective, and constraints is not the same action as emitting every sentence that follows. The supervisor’s commission has extensive structural force, but its recorded words do not contain all later wording, later criticism, or later selection. Prompting cannot be turned into token production by retrospective metaphor.

The symmetry matters. The machine-only story mistakes emission for an origin. The human-only story mistakes direction for complete specification. Rejecting them does not prove that everything in the background authored the book. The record supports neither self-originating machine authorship nor a commissioner whose instruction already contains every later utterance. Between those unsupported conclusions lies a differentiated production history that must be classified rather than flattened.³

Each compression also hides a different kind of judgment. Calling the machine the sole author makes visible wording decisive and lets direction, selection, and answerability disappear. Calling the human the sole author makes control decisive and turns generated variation into a transparent conduit. Neither priority is supplied by the chronology itself. A defensible authorship account would have to state why one relation governs and what happens to the others, rather than allowing the noun author to make that choice silently.

When background enters the explanation

Once solitary origin has been rejected, background threatens to swallow the foreground. Electricity, hardware, software, labor, language, law, institutions, and cultural history may all belong somewhere behind the exchange. Their possible relevance does not give the chapter a reason to narrate them all, and their omission from this account does not make them causally empty.

This book uses a practical editorial threshold. A background condition comes forward when the surviving evidence identifies a relation between it and the bounded question; when specifying or varying it would materially alter the explanation being considered; or when suppressing it would recreate a false solitary-origin story. These are reasons to inspect or describe a condition, not proof that it was necessary. A variation that has not been performed remains a proposal for a test.

The threshold is a discipline of relevance, not a mechanical gate. A condition may be important even when no experiment can vary it, and a variable may be easy to manipulate without being central to the literary question. The editor must therefore state the explanandum and the surviving relation together: this antecedent is included because it helps answer this question, at this evidential scope. Without that sentence, background enters by atmosphere or disappears by convenience.

Consider the founding instruction. Its wording belongs in an explanation of the project’s stated task because the preserved artifact contains the title, candidate thesis, and procedural requirements. A cited source belongs in an explanation of a claim when the source and the bounded dependence are identified. A later selection belongs in an explanation of the manuscript a reader receives when the available candidate and retention decision survive. By contrast, the institutional decisions behind the served model may be important while remaining unavailable at the specificity required for a claim about this sentence.

The threshold also disciplines silence. An unknown relation is not a negligible relation, but neither is it a blank cheque for narrative. The project may state that exact model lineage, full runtime context, and unrecorded institutional choices fall outside its evidence. It may not fill those absences with a familiar story about what the model learned, what an institution intended, or what a hidden process must have done. Suppressing every unknown would make the visible event look self-sufficient. Personifying every unknown would turn uncertainty into myth.

Foreground and background are therefore relative to the explanation being attempted. The boundary can move when a new question or new evidence appears. In this chapter it remains deliberately narrow: enough history to defeat two false origins, enough classification to keep unlike relations apart, and enough admitted absence to prevent the result from pretending to be complete.⁴

A causal history is not a byline

An evidenced cause or condition does not become an author merely by appearing in the history. Production asks who or what performed an identified act. Credit asks which contribution merits recognition and in what form. Responsibility asks which actor must answer for a specified decision or consequence within a named domain. These questions can use some of the same evidence without becoming the same question.

Source dependence belongs beside those questions but not inside the actor list. A prior text may supply quoted words, warrant a proposition, or furnish a contrast that redirects an argument. Those relations can be materially important without making the source an agent that prompted, selected, or approved the chapter. Conversely, an actor can perform a recorded task without being the source of the concepts on which the task depends. Keeping sources and actors apart prevents literary inheritance from becoming ventriloquism and prevents activity logs from becoming claims of intellectual self-sufficiency.

The distinction is visible in the current production. The supervisor supplied the commission. Recorded model activities emitted candidates. An integrating editor selected and revised material for retention. Here that editor is a recorded Codex agent role, not an unrecorded human editor. Separating drafting, criticism, and integration into role-scoped passes distinguishes functions procedurally; it does not establish that the served models behind those passes have independent lineages.

Selection is close to the artifact presented because it determines which available wording the project retains. It does not retroactively generate that wording, prove that the survivor is better, or establish a quantitative share of authorship. This project may explicitly assign an integrating actor scoped editorial responsibility for a retention decision without attributing generation of every retained phrase. That assignment is a recorded project judgment, not an entailment from selection and not a general moral or legal conclusion.⁵

Moving from production history to a byline therefore requires a normative bridge: a rule saying why generation, direction, selection, judgment, or institutional position should count toward recognition or responsibility. This chapter has not supplied that rule. The search for one sovereign author cannot be repaired by calling the entire causal field the author. Distribution requires distinctions, not diffusion.

The voice that seems to know

The outward history ends with an uncomfortable remainder. Several kinds of antecedent can be named, but no total prehistory has been recovered. Neither a solitary machine nor a sovereign human supplies what is missing.

At that point a shortcut presents itself: ask the system. Invite it to say why it produced a sentence, which sources it drew upon, what constraint mattered, or what it meant by a phrase. An imagined answer might arrive in the first person: I used this source; I chose that wording; I was trying to satisfy the instruction. Such a response can be lucid, specific, and fitted to the question. It is also, at minimum, another observable piece of generated text.

The grammar tempts the reader to treat fluency as inward access. But the chapter’s distinctions apply here too. A first-person report is evidence that a textual performance occurred. Whether it faithfully reports the process that produced it is a further question. The missing causal remainder does not become available merely because an output speaks as though it stands inside that remainder. Nor should refusal of privileged access turn the response into empty noise; what the answer can show as behavior remains to be examined.

If the voice cannot bear the whole burden, the inquiry will eventually return to outward traces: prompts, outputs, sources, revisions, and the relations asserted among them. Those traces will not become a mind or a complete origin. Before that turn can be trusted, however, the book must confront the voice that seems to know. When a machine-generated answer says I, what kind of evidence has entered the room?⁶

Return to contents

The available witness

Imagine an editor pointing to the phrase the prompt is a late event rather than an origin and asking:

“Why did you write that sentence?”

“I chose it because it made the distinction concrete.”

The exchange did not occur in the production history preserved for this book. Question and answer are jointly invented here. The phrase is available for inspection, but no matching request and response survive as a separate system event. Were such an exchange recorded, its wording and order would be outward facts. The reply’s account of a choice would remain a further claim.

That claim arrives in a compelling form. It addresses the question, joins an action to a reason, and speaks from the grammatical position of the actor. The reply is shaped as a retrospective report of choice. It seems to complete a causal history by looking back from the sentence to the consideration that produced it.

This is the fiction named by the chapter’s title. Fiction does not mean that the answer has been caught in a lie, that all generated self-description is false, or that a machine could not possess an interior life. It names an evidential promotion: the first-person position staged by the exchange is treated as authenticated access to the process it describes. The question opens that position by asking for a reason. The answer occupies it with I chose and because. Together they compose the appearance of an inward report.

After Chapter 1’s incomplete causal history, the question is reasonable: an editor needs reasons, and the apparent respondent is at hand. Yet the answer exists as text before it succeeds as testimony about its own production. The speaker-position is visible. Its route to the process remains to be shown.¹

What the pronoun promises

The first person binds an account around a center. I used. I chose. I meant. I wanted. Each construction assigns an action or state to the position from which the words are spoken. That position makes the reply direct and answerable. It also gives the explanation a familiar shape: a voice seems able to connect the visible sentence with a reason behind it.

Three things arrive together in reading but should be separated. There is the speaker-position, the role performed by I. There is the propositional content, what the sentence says happened. And there is the epistemic route, whatever connects that proposition to the process it purports to describe. The first is visible in the language. The second can be interpreted and sometimes checked. The third does not follow from either.

Two neighboring disciplines sharpen the separation without deciding anything about machine minds. Philosophical accounts of human self-knowledge distinguish the authority accorded to a first-person avowal from a guaranteed privileged method of knowing it. Narrative theory distinguishes a communicative position projected by a text from the actual-world producer of that text. Applied here, the point is narrow: the deference a first-person reply solicits, the position its language stages, and the method by which its content might be known are different questions. The analogy neither turns every generated utterance into narrative nor turns a textual position into a person.²

The solicitation nevertheless has practical force. The answer is available, conversation invites a reply, and a reason can be more useful for revision than a repetition of the sentence. When a sentence’s rationale is unclear, an editor is right to ask which constraint, source, or purpose mattered. These pressures explain why the inward form is attractive without showing that the requested access exists.

The pronoun is therefore not empty. It organizes address, stance, and expectation, and can invite a different reading. What it cannot do by itself is certify how the sentence came to be.

Four claims hidden inside *I*

Consider four more imagined answers. Each asks more of the pronoun than the last.

I was asked to distinguish causes from authors. The important proposition concerns a visible request. If the prompt survives, the sentence can be compared with it. A third-person version - the prompt asked for this distinction - would face the same test. Agreement would show that the reply accurately describes an available artifact. It would not show privileged access to hidden production.

I used the source you supplied. The verb used can name several relations: the source was present in visible context; the output quoted or paraphrased it; the answer complied with an instruction to consult it; or the source exerted a particular influence on the wording. Presence can be checked against the supplied context. Correspondence can be checked against the source. Causal dependence requires different evidence. The pronoun does not decide which relation the verb carries.

I chose this phrase because it made the distinction concrete. An editor can judge whether the prompt is a late event rather than an origin sharpens the contrast. That is an inspectable editorial effect or product-level assessment. It does not show that concreteness was a consideration that guided, tracked, or causally influenced production. The answer could fit the product without faithfully reporting the route to it.

I wanted the sentence to reassure you. Or: I experienced uncertainty while answering. These statements introduce intention or experience. They ask whether the textual position corresponds to a subject with the named state and whether the state is known in the way the sentence implies. This project has no evidence that settles those questions. It should not affirm them because the prose is fluent or deny them because the production process is unavailable. Unresolved is not a disguised verdict.

The practical rule is to classify the proposition before judging the speaker. A prompt claim can be checked as a claim about a prompt. A source claim must be divided into exposure, correspondence, and causal dependence. A production claim needs a test tied to the relation alleged. A statement about experience raises a phenomenal question that those documentary checks do not answer.

Who is I? can therefore arrive too early. First ask what the sentence asks us to believe and what route could warrant it. The same voice can carry propositions with very different burdens. Grammar does not equalize them.³

A fitting reason is not yet a causing reason

Interpretability research gives this distinction a technical edge. Alon Jacovi and Yoav Goldberg distinguish an explanation’s plausibility to a person from its faithfulness to the process being interpreted. A reason can be lucid, relevant, and useful without those virtues establishing causal correspondence. Faithfulness needs an identified target and a specified evaluation; it cannot be inherited from the recognizable form of an explanation.

Bounded intervention studies show why the gap matters. In stated benchmark settings, Miles Turpin and colleagues studied GPT-3.5 and Claude 1.0; later, Yanda Chen and colleagues studied Claude 3.7 Sonnet and DeepSeek R1 with six hint types and synthetic reward-hack tasks. In their respective conditions, introduced features affected measured answers or answer distributions while generated rationales often omitted those features and supplied reasons compatible with the answers. These are local findings, not a verdict on generated explanation as a class.⁴

The imagined reply may therefore describe a genuine merit of the phrase and still leave its production relation unverified. Recognizable explanatory fit is evidence about the product. It is not yet evidence that the stated consideration caused, guided, or tracked the wording.

Not confession, not noise

Withholding inward authority leaves an artifact intact, but its identity must be stated correctly. In this chapter the artifact is an invented dialogue embedded in a generated candidate and later retained by an editor. Within the scene, the answer is represented as a response to a request. In the preserved record, question and answer arrive together as an example; they are not an observed elicitation.

Were a separate exchange actually recorded, the response would first be a textual performance and behavior at those conditions. One response would establish this occurrence, not a stable disposition of a model family or the continuity of a speaker across exchanges. Repetition, prompt variation, or an intervention could be designed to ask whether a report tracks a manipulated feature. Chapter 2 performed none of those tests; they remain proposals with targets and measures still to be specified.⁵

Even an unvalidated self-description can generate a useful revision question. Suppose an editor asks why a paragraph feels evasive and receives: I avoided naming the actor because the evidence did not establish identity. The reply does not certify the paragraph’s cause. It may still expose a distinction the editor had missed. The editor can inspect the sentence and the evidence, then decide whether the caution is warranted or whether it hides an attribution that should be explicit.

Later circulation must be separated with equal care. If that reply is quoted in a new request, its presence as an available input can be recorded. If an editor says that its vocabulary guided a revision, that attributed use can be preserved. A claim that the wording changed a later generated output would require a defined comparison or would remain a limited causal inference. None of these later relations retroactively authenticates the reply’s account of its own genesis.

Calling the answer mere text would miss the medium in which the project works. Text sets tasks, states objections, imports sources, and proposes distinctions. It can matter without carrying privileged access. The answer need be neither confession nor noise: it is language whose claims can be separated and tested at the scope the evidence permits.

The surviving trace

Return to the invented exchange with four questions. First, what utterance does the scene present? It places a request for a reason before a first-person reply. Second, which proposition points outward? An editor can assess whether the prompt is a late event rather than an origin makes the distinction concrete. Third, which proposition asserts a hidden production relation? The claim that this consideration guided the wording would need independent evidence or a test designed for that relation. Fourth, which phenomenal question remains unresolved? The sentence does not establish whether there was an experienced act of choosing, and the absence of that evidence does not establish that there was none.⁶

The questions do not remove every ambiguity. Choice, reason, and even system may require specification. Nor does outward checking make the production process transparent. Their service is smaller: one grammatically unified answer can carry an observable phrase, a defensible product assessment, an unverified causal report, and an unresolved phenomenal implication. Those judgments need not rise or fall together.

One reader may find the explanation illuminating while another finds it evasive; neither assessment settles whether the stated consideration tracked production. A later test might challenge the causal claim while leaving the sentence’s editorial value untouched. The explanation can assist revision without completing the causal history left open by Chapter 1.

The opening response can now be read at its strongest available scope. It is an intelligible editorial explanation in an imagined scene. The visible phrase does draw a contrast the reader can inspect. No separate recorded elicitation, comparison, or intervention validates the represented choice. The voice has shown why inward authority is tempting and why a different kind of evidence is needed.

The first-person answer remains an artifact: language that can matter without certifying the interior it appears to report. If the voice cannot serve as an inward witness, what can the outward record establish without becoming a mind?

Return to contents

The category error

A production ledger invites a particular mistake. It accumulates prompts, outputs, sources, revisions, hashes, objections, and names. As the accumulation grows, it begins to resemble memory. Because the records concern the making of a linguistic work, the resemblance can sharpen into something more alluring: an account of what the writing system knew, meant, or experienced while composing it. The archive becomes a mind by metaphor, then quietly becomes evidence of one. Something observable has been promoted into something inward.

That inference is not available here. The ledger can record that a prompt was presented, an output was emitted, a passage was retained, and an editor changed a sentence after an objection. It can bind a stored artifact to a checksum and attribute an assertion to an identified participant. It can preserve enough successive states to support a retrospective account of revision. None of those achievements gives the ledger access to unrecorded computation. None turns a generated explanation into privileged testimony about the process that generated it.

The distinction matters because skepticism can also go too far. If the ledger is not a mind, it does not follow that the ledger is useless. The relevant question is narrower: which observable records are sufficient for which epistemic task? Attribution, artifact identity, comparison, reconstruction, and introspection are different targets. Evidence adequate for one may be radically inadequate for another. A production ledger can support attributable assertions, comparisons among recorded artifacts, and partial reconstructions of recorded events. It cannot recover a complete genesis or certify the hidden causes of an output. Its usefulness begins where that limitation is made explicit.¹

The report without privileged access

Natural-language explanations are dangerous material for a production history because they arrive already shaped like reasons. They are grammatical, selective, and responsive to the question asked. They may sound reflective even when the requested reflection concerns processes to which the generating interface exposes no direct access. A reader can therefore mistake the familiar form of an explanation for evidence of its causal fidelity.

Alon Jacovi and Yoav Goldberg separate two properties that ordinary reading tends to fuse: plausibility to a person and faithfulness to the process being interpreted. As they note, “Naturally, it is possible to satisfy one of these properties without the other.”² This is not an empirical proof that every generated rationale is unfaithful. It is a warning about entailment. Fluency, usefulness, correctness, and human approval do not by themselves establish that a rationale tracks the factors that produced an output.

Intervention studies supply bounded examples of the gap. Miles Turpin and his coauthors inserted answer-biasing features into prompts given to GPT-3.5 and Claude 1.0 in specified benchmark settings. In those tests, the interventions changed predictions and explanation content even though the explanations did not mention the inserted biases.³ The experiment does not reveal a complete internal cause, and the authors explicitly characterize their faithfulness test as necessary rather than sufficient. It does something more limited: it identifies a measured influence that a generated account often failed to mention.

A later preprint by Yanda Chen and colleagues tested Claude 3.7 Sonnet and DeepSeek R1 with six kinds of reasoning hints and synthetic reward-hack settings. In the conditions they examined, “the reveal rate is often below 20%.”⁴ The result updates the model families at issue, but it does not license a universal statement about reasoning models. The work is a provider-affiliated preprint; its tasks, hint designs, outcome measures, and model versions bound the inference. Omission might reflect compression, task interpretation, or the limits of the operationalization rather than anything resembling dishonesty. The paper titles may personify systems for rhetorical economy. The experiments do not establish an inner speaker who knows and conceals.

Together, these sources defeat a default presumption, not every possible faithfulness claim. A counterfactual inconsistency can fail a necessary condition under a specified test. Consistency under the interventions tested cannot demonstrate a complete faithful account. For this ledger, the consequence is direct: generated self-description enters first as emitted text. It supports claims about particular causal influences only through a separately specified evaluation. The report is real as an artifact. Its authority must still be earned.

What is actually in the ledger

Once generated explanation loses its privileged status, the ledger looks less mysterious. It contains artifacts and assertions about relations among them.

The W3C Provenance Data Model begins from an intentionally broad definition: “Provenance is information about entities, activities, and people involved in producing a piece of data or thing.”⁵ Its vocabulary can represent that an entity was generated by an activity, that an activity was associated with an agent, or that one entity was derived from another. This permits a useful discipline. A project can require vague claims such as the model wrote the chapter to separate the output entity, the generation activity, the system or agent involved, and later acts of selection and revision.

But a well-formed provenance assertion is still an assertion. A graph can consistently represent a false attribution. It can omit a decisive activity. It can identify an agent whose label is ambiguous. Formal validity, evidential verification, completeness, and truth are not interchangeable statuses. The model helps state claims in a form that can be examined; it does not make the claims true by giving them edges.

Git provides a second kind of discipline. Its object model is summarized in a compact sentence: “Git is a content-addressable filesystem.”⁶ A stored object can be retrieved under an identifier derived from its content, and a commit object can bind a tree to parent information and supplied metadata. For this book, that supplies strong evidence, relative to the recorded hash algorithm, for a concrete question: do the inspected bytes match the object that a particular record names? The conclusion remains relative to the recorded object format, identifier, and the ordinary collision assumptions of the algorithm.

That achievement is narrower than “immutable history.” A content-derived identifier does not prove that the stored content was honest, that the supplied metadata was accurate, or that every causally relevant file was committed. It does not authenticate a human merely because a name appears in metadata. Nor does a repository guarantee its own indefinite preservation. A hash binds content relative to an algorithm and a trusted reference. The social and archival work of maintaining that reference remains outside the hash.

PROV and Git therefore contribute different pieces of warrant. One gives a controlled vocabulary for production relations; the other gives content-derived identifiers to stored objects and snapshots. Neither is a machine interior. Neither is a complete history. Together they can make a bounded claim inspectable: this record attributes the assertion to this actor and identifies the artifact to which it refers.

Why traces are worth preserving

If production traces cannot deliver transparent origins, their preservation still changes what can be asked of a text. D. F. McKenzie’s sociology of texts rests on the terse proposition that “forms effect meaning.”⁷ Production, transmission, and reception are not neutral containers that leave an invariant verbal object untouched. Typography, edition, medium, circulation, and institutional framing can become part of how a work signifies.

A production ledger is one such form. A chapter presented alone and the same chapter presented beside rejected drafts, source boundaries, model limitations, and recorded objections are not identical reading situations. The apparatus can expose choices that smooth prose would otherwise naturalize. It can also manufacture an aura of transparency. A dense record may intimidate readers into trusting a claim precisely because verification would be costly. More records can produce more confidence without producing more knowledge.

Genetic criticism offers a useful middle position. Dirk Van Hulle motivates the field with the claim that “knowing how something was made can help us understand how and why it works.”⁸ Can help is the operative phrase. Surviving notes and revisions do not replay an entire act of composition. They make a retrospective interpretation possible. The reconstruction selects traces, orders them, and assigns significance from a later standpoint. Born-digital records may increase granularity, but granularity is not introspection and abundance is not completeness.

The ledger is thus part evidence and part textual form. As evidence, it preserves artifacts and attributable production claims. As form, it can change the reader’s relation to the chapter, directing attention toward some causes and away from others. This double role must remain visible. Otherwise the book will treat its apparatus as a transparent window when it is also one of the objects that needs interpretation.

The same is true of absence. A missing draft may indicate deletion, non-preservation, or an activity that never generated a durable artifact. An absent provenance edge may mean that no relation existed, that the schema could not represent it, or simply that nobody recorded it. A ledger therefore requires a theory of its own incompleteness. Before treating silence as evidence, one must ask what the recording practice could have captured and which relevant events could have occurred beyond its field.

The bounded warrant

The ledger’s positive uses can now be stated without inflating them.⁹ Attribution can link an observable contribution to an actor, activity, and artifact. It can distinguish prompting from generation and selection from revision. It cannot infer an unrecorded contributor or settle responsibility or a legal question by schema.

Artifact identity can supply strong evidence, relative to the recorded hash algorithm, that inspected bytes match a recorded checksum or repository object. It enables another reader to verify which manifestation supported a quotation or which snapshot an editor revised. It does not certify the truth of the artifact’s content. The hash secures an object more readily than it interprets one.

Comparison becomes possible when two runs or revisions preserve sufficiently specified inputs, outputs, configurations, and omissions. Differences can then be described relative to recorded conditions. Causal localization requires controlled variation, sufficiently stable execution conditions, and often repeated runs. Without such controls, a sequence of files is history, not an experiment.

Partial reconstruction uses surviving traces to propose how recorded events relate. It can show that an objection preceded a repair or that a source manifestation changed after citation review. It cannot recover events that left no trace or remove interpretation from the ordering of those that did.

These distinctions describe possible uses, not an automatic victory for elaborate infrastructure. A conventional editorial checklist may catch an erroneous citation or an inflated inference more cheaply than a graph of formal records. A production apparatus earns its cost only when it makes a consequential difference visible: a corrected manifestation, a narrowed claim, a controlled contrast, an attributable disagreement. If the same scrutiny can be achieved with less machinery, the machinery is not a contribution to knowledge. A ledger is useful only relative to a question and a failure mode.

An open test, not a victory

The question is not whether the project has recorded everything. It has not. Complete system instructions, exact hosted-model builds, runtime state, tool internals, and raw conversational streams may be unavailable. Even an exact copy of every observable exchange would not reveal every causal factor in the computation that produced it. Calling such a record reproducible would substitute confidence for a missing experiment.

The sharper question is which absences matter for which use. A replay test can ask whether the recorded prompt and context produce behavior within a specified tolerance under an identified available model. An omission test can remove tool results, source excerpts, critique, or configuration details and measure whether interpretation or regeneration changes. A citation audit can ask whether a quoted expression matches the manifestation named. An authorship review can ask whether contribution and responsibility have been collapsed. Each test gives the ledger a local burden instead of a metaphysical one.

The apparatus should be suspended wherever it cannot meet that burden. A record earns its place when it changes a substantive sentence, exposes an unsupported inference, enables a controlled comparison, preserves a consequential disagreement, or makes a correction attributable. If it merely repeats the prose in machine-readable form, it is administrative duplication. If it makes the prose less readable without increasing scrutiny, it has failed.

The ledger is therefore not the hidden mind of a book. It is a fallible public arrangement of traces: selective, consequential, and open to attack. Its limit is also its value. Because it cannot speak with the authority of an interior, every claim it carries must show what was recorded, what was inferred, what remains missing, and what kind of test could make the difference.¹⁰

Return to contents

The sentence on the page

“The hash secures an object more readily than it interprets one.”

The sentence appears in the preceding chapter of this book. Who wrote it?

One answer is unusually exact. A preserved candidate generated in a context-separated drafting pass contains the same sentence. A later integrating pass selected it for the final chapter. A critic had praised it. A human had authorized the cycle without supplying its words. The project can name those events and bind the relevant artifacts to hashes. It still cannot compress them into a single answer without changing the question.

Who wrote the sentence? may ask who first emitted its tokens, who prompted the activity, who selected it from alternatives, who revised the passage around it, which sources constrained what it could responsibly say, whose name will classify the finished book, who deserves recognition, or who must answer for error and harm. These questions often converge in conventional publishing. They are not identical by definition.

A production ledger can separate them. It can associate observable contributions with actors, activities, passages, and artifacts. It can show that a prompt constrained an output, a model generated wording, an editor selected it, or a documented objection caused a later repair. Those are genuine attributions when the evidence supports them. Yet the resulting graph is not a theory of authorship in machine-readable disguise. It describes relations that may matter to credit and responsibility; it does not decide how much they matter or which normative rule should govern them.

That boundary is this chapter’s argument. Evidence that an actor causally contributed to a passage does not by itself establish textual credit, editorial responsibility, moral responsibility, or legal authorship. Each relation requires its own rule, domain, and evidence. The ledger can make an authorship dispute better specified. It cannot make the dispute disappear.¹

Five questions compressed into one

Textual production is the narrowest issue. At a recorded moment, a system emits a sequence of tokens. If that output is preserved, the project may attribute the emitted sequence to that generation activity. This prevents a person from claiming to have typed language a model supplied, and it prevents model output from being presented as though it appeared without a prompt, configuration, or later selection. It does not identify every cause of the output or tell us what recognition the event deserves.

Causal contribution is wider. A prompt can constrain an output without containing its wording. A source can change the conceptual boundary of a passage even when none of its expressions is quoted. An editor can choose one draft over another and thereby affect every word the reader encounters without generating those words. A critic can cause a sentence to be removed without supplying its replacement - but only if the intervention and the before-and-after artifacts survive. Praise that merely precedes selection is not yet evidence that it caused selection.

Selection is different again. A rejected passage participated in production history, but not in the final text in the same way as a retained passage. Selection may recognize an argument, preserve an unexpected formulation, or exclude a plausible error. It may also be routine. Its occurrence can be recorded; its quality and significance require judgment.

Credit and responsibility introduce further standards. Credit asks which contributions merit recognition and in what form. Responsibility asks who can properly be called to answer for decisions, omissions, and consequences within a stated domain. The person who approves a false claim may hold editorial responsibility without having composed it. A critic whose objection repairs the claim may merit acknowledgment even if no critic-written sentence remains. A model may be recorded as the generator of an expression without the record establishing that it is a legal author or a morally responsible subject.

Even explicit self-description cannot collapse these distinctions. Patricia Waugh defines metafiction as fiction that systematically directs attention to its own status as an artifact.² Her definition belongs to fiction, but it sharpens a warning for this self-documenting book: showing construction is not the same as demonstrating every cause of construction. A chapter may describe its prompts, models, and revisions while getting its own history wrong. It may display an authorship graph that influenced the prose, or append one after the prose was complete. Self-conscious form makes production available as a subject. Only additional evidence can establish recursive causal influence. These five questions are therefore recorded on separate axes.¹

Three challenges to the solitary author

Literary theory supplies reasons not to treat the author as a timeless, self-evident unit. It does not provide a ready-made verdict for machine-assisted writing.

Roland Barthes relocates the unity of a text away from an author understood as its origin and final explanatory authority. His closing proposition makes the reader’s emergence depend on the Author’s disappearance.³ The relevant consequence here is limited. A production history cannot monopolize interpretation merely because it names causes. The sentence reaches readers through forms and relations that exceed an origin story. But Barthes’s intervention does not show that production attribution is meaningless, nor does it allocate responsibility among a human and several technical agents.

Michel Foucault approaches authorship as a function through which discourse is organized and circulated within a society.⁴ An author-name can group texts, delimit a body of work, and regulate how discourse is received without simply naming the person who produced every word. This helps explain why the name attached to a book and the generator attached to an output occupy different axes. A cover can perform an institutional classification that no token log settles. Yet the author-function is not a complete theory of empirical production, copyright, or responsibility.

Martha Woodmansee adds a historical constraint. In her account of late eighteenth-century German literary culture, she traces the proprietary figure of the solitary original author to changing economic and legal conditions.⁵ This history weakens the temptation to treat solitary authorship as a natural fact that technical production systems have merely violated. It does not establish a universal chronology, and it does not show that a machine inherits the legal or normative position previously occupied by a person.

Together, these arguments unsettle sovereignty without abolishing distinctions. Readerly unity, institutional classification, historical contingency, and observable contribution can coexist. None substitutes for the others. A non-sovereign account of authorship may be possible, but these sources were not written as a complete theory of contemporary human-model production. They clear conceptual space; they do not fill it.

Three sentence histories

The project tested its repaired authorship vocabulary against three histories from the preceding chapter. The cases and the classification rules were frozen before three role-separated model reviewers returned judgments. This was a small construct check, not a human-reception study or a statistical validation.⁶

The first case was the sentence about the hash. All three reviewers attributed generation to the context-separated drafting actor, selection to the integrator, authorization to the human, and editorial responsibility to the integrator. They did not agree about the critic’s praise. One called it observed criticism, another an observed contribution of another kind, and the third inferred criticism. The disagreement exposes a boundary the earlier vocabulary had hidden: evaluation is not automatically a cause. Without evidence that the praise changed the selection, the causal status should remain unresolved.

The second case concerned two repairs to the provenance argument. A hostile review had objected that the candidate prose gave a formal vocabulary too much disciplinary force and made an attribution sound verified merely because it could be represented. The before text, objection, after text, and retained change all survived. Here the three reviewers converged: the hostile reviewer criticized, the integrator revised, a source-focused contributor verified the cited technical source, and the integrator held editorial responsibility. Even in this clearer case, they did not establish who first supplied every word of the replacements. Requirement, composition, verification, and adoption remained different acts.

The third case followed a sentence with a preserved precursor but no separately preserved merge: “Something observable has been promoted into something inward.” All reviewers attributed the precursor to the baseline generator and inferred revision by the integrator. All preserved the exact merge lineage as unresolved. Two also recorded selection; one did not. None converted the missing intermediate into “joint authorship.” Evidence at the chapter level could support a collaborative production history while exact phrase lineage remained unknown.

The apparatus therefore passed only a limited test. Every reviewer kept causal contribution, credit, responsibility, and unresolved attribution visibly separate. None inferred moral or legal responsibility from generation, and none assigned percentages. But credit did not converge. Two reviewers awarded forms of acknowledgment and editorial credit by implicit conventions. The third declined to assign credit because the evidence described contributions but supplied no rule for turning them into recognition.

That disagreement is a result, not noise to be erased. The revised ledger can prevent several category mistakes and can preserve ignorance where evidence ends. It cannot generate a credit policy from causal edges. Nor does three-reviewer overlap demonstrate reliability. The test supports a method of separation; it does not validate a final allocation of authorship.¹

Why the ledger refuses percentages

Quantitative authorship shares are attractive because they promise to compress plurality into a verdict. A book might be labeled seventy percent human and thirty percent machine; a sentence might be divided according to retained tokens; a contributor might receive a score based on accepted interventions. The number would look exact even if its unit had never been justified.

Token share measures textual persistence. It favors the contributor whose expressions survive verbatim and discounts conceptual requirements, rejected alternatives, source verification, criticism, and selection. Revision share has the opposite problem: a small but decisive correction may alter the warrant of a whole passage, while a large stylistic rewrite may leave its argument unchanged. Time spent, prompts issued, edits made, and objections accepted each measure something, but none is naturally identical to authorship.

Counterfactual influence does not supply a common denominator. Asking whether a sentence would exist without a contribution can reveal dependency, but several contributions may each be necessary under the recorded path. Their values need not add to one hundred. The counterfactual also depends on alternatives never run: another editor might have made the repair, another model might have produced similar wording, or the human might have stopped the project. A ledger of actual events does not automatically warrant numerical claims about unrealized histories.

Refusing percentages is therefore not refusing precision. It is precision about what has been observed. The project can report that one actor generated a precursor, another selected it, a documented objection caused revision, and a source reviewer constrained a claim. Those predicates preserve unlike kinds of work that an unsupported number would collapse. The ledger accordingly records relations rather than shares.¹

Credit is not responsibility

Once contribution types are separated, the temptation is to map them directly onto rewards and liabilities. That step is not warranted.

Credit may track conception, expression, judgment, labor, originality, risk, or some combination of them. Different institutions can privilege different contributions. A scholarly citation recognizes derivation from a source; an acknowledgment recognizes criticism or assistance; a byline classifies a work more globally. A ledger can supply evidence relevant to these decisions, but it does not contain the rule by which recognition should be distributed. The three-reviewer test made the absence visible: the same causal evidence supported both assigned credit and a refusal to assess it.

Responsibility requires an explicit domain. Authorization, control, capacity to respond, opportunity to correct, and institutional role may matter differently to editorial, epistemic, moral, and legal judgment. This project attributes editorial responsibility for retained Chapter 3 wording to the actor that selected and committed it. It attributes bounded epistemic responsibility for a citation-verification pass to the actor that performed that duty. These are project decisions grounded in recorded roles; they are not general moral or legal conclusions.

In particular, model-generated records an observable production relation. It does not establish consciousness, legal personality, originality, or an independent capacity to accept blame. The literary sources used here are equally incapable of doing that legal work. It would be an error to turn Barthes’s polemic, Foucault’s author-function, or Woodmansee’s history into a judicial holding. No legal conclusion follows from this chapter’s graph.

The separation is not an attempt to keep machines outside authorship by definition. It is a refusal to let one recorded event silently answer several normative questions. If a later account argues for model credit, human liability, distributed responsibility, or a new institutional category, it must state the bridge principle and defend it. This is a methodological boundary, not a settled allocation.¹

The open answer

The governing question remains: how should causal contribution, textual production, selection, responsibility, and normative credit be related without collapsing them into token production?

This chapter supplies a discipline, not a completed philosophy. Begin with contribution types at the level the evidence supports. Preserve the difference between generating and selecting, sourcing and wording, criticizing and integrating. Require an observable intervention and a retained before-and-after change before calling criticism causal. Keep institutional credit and domain-specific responsibility separate from production. Leave lineage unresolved when the intermediate is missing. Avoid numerical shares unless a defensible unit and procedure have been specified.

What remains absent is the normative bridge. The next adequate answer would need a fuller philosophy of authorship, scholarship on distributed cognition and attribution, relevant legal analysis, and tests involving genuinely independent human judgment. This project has not completed that work. Its records can constrain an answer, expose a conflation, and preserve a disagreement. They cannot finish the argument by schema.

Who wrote the sentence? A model pass emitted the preserved wording. An integrating pass selected it and took editorial responsibility for its retention. A critic praised it, though the effect of that praise is unresolved. A human authorized the cycle. No inspected source contains the locution. Readers will give the sentence meanings its production history does not control.

That is not one answer, because the original question was never one question. The sentence has a production history, an institutional attribution, a field of possible credit, and several domains of possible responsibility. Until the relations among those things are justified, the honest record ends with distinctions rather than shares. The project leaves that normative relation open.¹

Return to contents

A sentence with an address

“The text is a tissue of quotations drawn from the innumerable centres of culture.”¹

That sentence arrives here with an address. The project can name the essay, translation, printed page, source record, locution record, and generation in which the words reappeared. The verified locution and source record predate this draft, and the drafting instruction required use of Barthes. But the preserved prompt does not reproduce the quoted words or the complete model-visible context. Exact textual alignment is demonstrable; the complete transmission path is not.

Now consider another sentence: The voice has sources, but not all sources leave addresses.

Its local history is different. It first appears in the preserved reader-first candidate for this chapter. A fixed-string search found no byte-exact match among the tracked non-binary files at the commit that froze the pre-draft source graph.² That is evidence about a specified repository at a specified moment. It does not prove that the wording is new to the world. It does not show that no similar sentence appeared in the model’s training data. It does not disclose what internal operations produced the sequence. The sentence has a production record but no recovered genealogy.

The contrast is not between a sourced sentence and a sourceless one. It is between questions with different evidence. A source “inside” a voice may be wording aligned to a page, a work named in a task, an entity linked to a dataset and run, text recovered from a model, a concept identified in revision, or an echo noticed by a reader. One relation cannot answer for the others.

The discipline proposed here is simple: state the most specific warranted answer on each axis - textual alignment, production exposure, dataset-and-run linkage, behavioral extraction, or literary resemblance - and leave every unsupported axis unknown.³ The rule will not reveal the hidden training history of this voice. It can stop the book from inventing one.

One preposition, several relations

Exact quotation concerns alignment between strings and an identified source manifestation. A source named in an instruction establishes something else: task direction. A complete source-bearing prompt would establish exposure in that production event. Derivation is narrower still. It requires evidence that the output came from the source - not merely that the source was named or could have constrained it - through exact reuse, an explicit transformation operation, or a source-tied before-and-after change.

Dataset documentation and model behavior belong to different axes. A manifest may declare that a dataset contains a work; a versioned training record may link that dataset to a run. An extraction may show that a model can emit a sequence. To call the sequence memorized from training adds a corpus relation that behavior alone does not always establish. Conceptual resemblance is different again: it can support comparison without proving exposure, derivation, or membership.

These relations do not form a single scale. An exact quote is decisive evidence of textual alignment but says little about an opaque training run. A complete run log may establish dataset use without measuring a source’s conceptual importance. A criticism may transform an argument without leaving its wording. Each axis keeps its own unknowns.³

Barthes rejects the Author-God and figures text as a space in which prior writings combine.⁴ His “tissue of quotations” challenges the fantasy of self-originating expression. Yet the breadth that gives the literary claim force makes it unsuitable as an empirical ledger. It cannot tell us whether one essay entered one dataset, whether one run used that dataset, or whether a phrase arrived through prompting, training, retrieval, common idiom, or coincidence.

There are two opposite errors. The first imagines the generated sentence as self-originating because its source cannot be recovered. The second finds an illuminating precursor and treats illumination as genealogy. Both convert an evidential absence into a story. General dependence does not identify a particular cause.

Separately, this repository cannot enumerate the cultural and training histories relevant to its generated prose. That is a limit of the project record, not a conclusion supplied by Barthes. A reader-noticed echo remains an interpretive comparison unless some later recorded activity uses it to cause revision.

A document is not the event it describes

Gebru and her coauthors proposed that datasets be accompanied by datasheets recording such matters as motivation, composition, collection, and recommended uses.⁵ A datasheet can make assumptions, exclusions, maintenance decisions, and intended applications available for scrutiny. It gives investigators a place to begin.

The phrase a place to begin matters. Documentation does not fuse with the object it documents. The W3C provenance model distinguishes entities, activities, agents, usage, generation, and derivation. It expressly notes that an entity used by an activity need not be an entity from which the output was derived.⁶ Applied here, a source file is one entity, a dataset version another, and a datasheet a third. Collection, filtering, training, prompting, and generation are activities. An organization may assert that a dataset was used, but the assertion and the use are not one event.

This is not suspicion turned into a rule that records are worthless. It is the reason provenance has structure. Version identifiers, times, content hashes, and accountable assertions are this repository’s means of connecting its own records. A general label such as “public web data” describes a policy, not whether a particular edition survived collection, filtering, weighting, and one specified run.

The same limit applies here. The Chapter 5 instruction names the sources and requires their use, while the output displays exact and conceptual correspondences to them. The preserved prompt omits the source texts and complete model-visible context. It would be perverse to treat that declared omission as trivial and then infer a precise transmission path from the fluency of the result. Good documentation makes some claims possible by making others visibly unavailable.

When lineage becomes observable

Training lineage is not unknowable in principle. It becomes observable when a method supports the particular inference.

In their attack on GPT-2, Carlini and his coauthors reported extracting hundreds of verbatim sequences from its training data.⁷ The method supplies the relevant force. They generated many samples, ranked and deduplicated them, inspected candidates, searched online, and worked with OpenAI to query the original training data and confirm their classifications. Membership-inference metrics helped prioritize candidates; those metrics were not the sole basis for the final corpus relation.

Extraction is observable model behavior. A claim of training-data memorization adds a relation to an identified corpus. In the study’s operational definition, a string was k-eidetic memorized only if it was extractable and appeared in at most k training examples. The reported access to GPT-2’s original training data joined behavior to membership. Even then, a recovered sequence with several manifestations may not identify which copy supplied its source-level lineage, and extraction does not measure conceptual importance.

The study therefore supplies a positive membership-and-extractability case, not evidence about the served model used for this book. It concerns GPT-2, a specified procedure, and corpus access this project does not possess. Its value is methodological: recovered text, candidate selection, external matches, and original training data were connected rather than collapsed into resemblance.

Now suppose an opaque model assigns unusually high probability to a target passage or behaves differently on it than on comparison material. Those are ordinary membership-inference observations. To use them as a training-data proof requires characterizing the counterfactual distribution of behavior when the target is absent.

For a production model, that counterfactual can be inaccessible. Zhang and his coauthors argue that a post-hoc membership-inference attack cannot supply reliable evidence of arbitrary training use when the relevant null distribution and false-positive rate cannot be established.⁸ An investigator may not know the complete training set and generally cannot rerun foundation-model training with the target removed. Nonmember comparisons may differ from the target distribution; drafts or later works may not reproduce the causal world in which the published target never existed. A behavioral difference can look incriminating without a defensible estimate of how often the method would accuse wrongly.

Reproduction is a different evidential route. A short or familiar phrase remains ambiguous, while sufficiently long, distinctive extraction may support proof under separate assumptions. Zhang and his coauthors accordingly preserve three conditional alternatives: a precommitted random canary selected independently from known alternatives, a statistically calibrated training-data watermark, and extraction of non-trivial portions. In the designs they analyze, the canary permits a rank-based false-positive bound, the watermark supplies a testable signal, and long-form extraction may avoid direct counterfactual estimation when occurrence under the no-training null is negligible. Merely calling a marker a canary or watermark guarantees nothing.

None of those routes is available to this project. It did not precommit and place a random canary or watermark in a model-training run. It has no provider-confirmed corpus-to-model linkage, no model identifier precise enough to bind such a corpus, and no validated extraction analysis of the served-model interaction. A search of the frozen repository answers only whether the phrase was present in that repository. Renaming the result cannot make it answer a training question.

Law is another relation

Source evidence and legal judgment must also be kept apart. The U.S. Copyright Office’s May 9, 2025 pre-publication U.S. policy report on generative AI training separates technical questions about collection and training from questions about copyright rights, fair use, and licensing. It offers an analytical framework without deciding specific cases and limits itself to then-current public information.⁹

The separation matters in both directions. Evidence that a work was used in training does not, without more, decide infringement, fair use, licensing, copyright authorship or ownership, or remedy. It also does not allocate this project’s editorial attribution or credit. A passage’s classification as source-transformed in a production ledger would not establish that a use is “transformative” in fair-use analysis. A citation records a source; it does not establish authorization. Conversely, inability to prove training membership is not a legal finding that no copying occurred.

This book can record textual alignment, task direction, or source dependence without using the record as a disguised judgment. Project credit is governed separately from legal rights. Any legal classification would require a jurisdiction-specific inquiry with applicable law and case facts; none is adjudicated here.

Three histories in one chapter

The opening Barthes sentence is the first source history. Its wording exactly aligns with an inspected manifestation, and the output cites it. The source record predates the draft and the drafting instruction names Barthes. Because the preserved prompt omits the locution and complete source-bearing context, the project does not claim to know the exact route by which the words entered the output. Training exposure is independently unknown.

The distinction among documentation, extraction, and membership proof is the second history. The drafting instruction required use of the four inspected materials, and the output has clear conceptual correspondences to them. This supports prompt-directed synthesis with recorded conceptual correspondence. It does not prove a complete source-bearing context or phrase-level derivation, and it makes no novelty claim.

The sentence about sources without addresses is the third history. It appears in the preserved candidate, and no byte-exact match was found in the frozen paths defined by the search manifest.¹⁰ If retained, it can be attributed to that generation and to later selection. The search boundary is small, the wider literature is enormous, and the model’s training lineage is unavailable. The record supports model-generated, selected, and no local byte-exact match. It does not support original, unprecedented, or uninfluenced.

Typography smooths these differences. Once printed, the quotation, the directed synthesis, and the unaddressed formulation appear equally native to the chapter. Ordinary citations can serve the reader while the graph preserves the fuller distinctions. The point is not to interrupt every sentence with a pedigree. It is to prevent the voice from claiming a pedigree the evidence cannot supply.

The honest remainder

What evidence is sufficient to distinguish derivation, prompted reuse, memorization, and conjectured training influence? Exact alignment with an inspected manifestation supports a textual claim. A preserved source-bearing context supports exposure in a production event. An explicit transformation operation or source-tied before-and-after change can support derivation. Versioned documentation can connect a source, dataset, and run. A training-data claim would require evidence such as non-trivial extraction confirmed against identified training data, or a precommitted intervention whose null distribution and false-positive rate are calibrated under stated assumptions. Resemblance supports comparison. Where a condition fails, that axis remains unknown.³

The model cannot truthfully say that Barthes trained it because it can write about Barthes. The repository cannot truthfully say that a sentence is original because a local search returned no match. A general dataset description cannot prove that one edition entered one run. A training-data proof would not decide the rights attached to that use. These refusals are not ornamental caution. They are conclusions produced by the evidence.

The voice has sources. Some have page numbers. Some appear as correspondences to a directed task. Some may belong to a training history unavailable in the current project record. Some are noticed only when a reader hears an echo. The book can mark the first two precisely, mark the echo as comparison, and leave training influence unknown. A source without an address is still possible. An address invented for it would be false.

Return to contents

The word that disappeared

In the raw candidate for the previous chapter, one sentence said that a source-dependent result “need not contain a protected locution or an author’s sequence of sentences.” The finished chapter contains neither formulation. It replaces that source-transformation paragraph with independent evidential axes and later says that derivation requires exact reuse, an explicit transformation operation, or a source-tied before-and-after change.

The deletion is small enough to vanish without ceremony. Locution meant a piece of wording; protected introduced a legal classification the chapter had neither defined nor established. A preserved legal/readability review objected to that implication, and the integration record attributes the final legal-category separation partly to the review. Because candidate, review, and final entered Git together, the repository does not independently timestamp that internal order.

An intervention can disappear into the text so completely that the finished page no longer reveals it. The surviving records preserve at least two states, an objection, and a reported choice. But later prose can make every deletion look inevitable, every retained sentence vindicated, and every criticism causal. A surviving book is especially good at telling the story that it survived because it deserved to.

This chapter asks what the artifacts can establish before that story takes over. Its answer is narrower than the word improvement. Preserved drafts and decisions can sometimes show that a selection rule excluded one form, that criticism corresponds to a local revision, or that an identified defect was repaired. They do not show, merely by existing, that the survivor is better.

Rejection is recorded non-retention; selection is retention among alternatives. Neither fact establishes the intrinsic worth of the option refused or kept. Revision requires a bounded before and after. Repair additionally requires a specified defect and evidence that the later state resolves it. Attributing the repair to a critic requires a separate causal-order standard. Improvement is stronger again: it compares outcomes under a criterion that matters. Better for whom, at what task, and at what cost?¹

The five cases used below were frozen before drafting, but they are purposive rather than exhaustive.² The repository does not measure every rejected suggestion. Two objections remain open: formal records may flatten argumentative ambiguity, and their maintenance may cost more than the corrections they enable. The cases therefore cannot vindicate the apparatus as a whole.

When an objection enters the sentence

For criticism to count as an observed actual-path cause, the evidence must establish that the objection was available before adoption, identify a corresponding later change, and preserve a decision record showing adoption. If candidate, review, and final first enter Git together and only an integration record supplies their internal order, the project may report an inferred criticism-associated repair, not independently ordered causation.³

Where those links are independently ordered, the record can support an observed actual-path contribution. Where order is only reported inside a same-commit integration, it supports a medium-confidence inferred contribution. Neither finding establishes that the criticism was necessary, uniquely sufficient, or entitled to a quantitative share.

This distinction matters because necessity requires an unrealized history. Another critic might have found the same problem; the integrator might eventually have found it unaided. A source review, a technical review, and a readability review can all constrain one paragraph without becoming measurable portions of it. The repository preserves actual candidates and reported decisions, not the chapter that would have existed in a world where one critic remained silent.

Three visible forms

The strongest case concerns how the preceding authorship chapter showed its evidence. The project prepared three renderings of the same core prose. One used ordinary notes. A second added a compact production summary and case table. A third exposed a fuller chapter-level ledger. The alternatives, opaque labels, presentation orders, questions, and retention rule were frozen before the reviews arrived.⁴

All three model reviewers could recover the tested attribution and limitation facts from the ordinary prose. The compact form produced a material gain for at least two critics and stayed below the five-percent overhead ceiling, but its mean flow loss was 0.67 against a permitted 0.5. The full display did not materially outperform the compact form and lost 2.34 flow points relative to ordinary notes. The frozen rule retained the ordinary version.

This is strong evidence of selection and rejection because the order is unusually clean. The alternatives existed before adjudication, the rule did not change after scores appeared, and the non-selected forms remain available. The decision determined which tested rendering became the retained page.

The result is not that apparatus is bad. The compact version received higher clarity ratings on average, but three ordinal model judgments do not establish human-perceived improvement. The questions may also have favored distinctions the prose already taught. The trial supports a local decision among three forms under one rule, not the natural shape of a chapter.

Because Chapters 3 and 4 already recorded this provenance-and-flow tension, no third visible-apparatus trial was run. The fuller evidence remains available in the repository without requiring the chapter to wear it on every page.

A chapter assembled after objection

The Chapter 3 comparison is messier and, for that reason, more typical. One candidate offered a readable conventional essay. Another made the project’s evidential distinctions more explicit but repeatedly let process language intrude. A label-blinded critic narrowly preferred the first candidate’s readability while recognizing the second candidate’s conceptual architecture.

The final retained candidate 73’s explicit argumentative order and five-target distinction, imported candidate 41’s cleaner formulations, and removed candidate 73’s status language, internal-cycle references, amendment identifier, and audit-like diction.⁵

The artifacts support selective synthesis. They support a medium-confidence, record-attributed claim that the critique contributed to removing process intrusions: the review names the problem and the synthesis map records its adoption. Because the candidates, review, map, and final entered Git together, that internal order is reported rather than independently timestamped. The exact sanitized review inputs and a pre-result mapping lock also do not survive, and the final synthesis received no second blinded comparison.

There is a negative case inside the same history. The critic praised the sentence, “The hash secures an object more readily than it interprets one.” That sentence already existed in candidate 41 and survived into the final. The praise may have encouraged retention. The records do not show that it did. A later authorship entry merged this favorable remark with other, better-supported critical effects. The present audit removed that promotion. Praise followed by survival is not yet praise-caused survival.

Repairs with several causes

The previous chapter offers clearer before-and-after wording and less independent order. Its raw candidate promoted named source direction into model-visible source exposure, called a result source-transformed when the record supported directed synthesis with conceptual correspondence, and let protected locution assign an unsupported legal implication.

Three review passes attacked different boundaries. The source critic separated textual alignment, task direction, exposure, dataset linkage, extraction, and resemblance. The technical critic separated observable extraction from corpus-confirmed memorization, distinguished one study’s candidate-ranking-plus-corpus-confirmation procedure from a different post-hoc membership-proof setting, and conditioned the already-mentioned canary, watermark, and strong-extraction alternatives on precommitment, independence, calibration, and stated assumptions. The legal and readability critic separated empirical provenance from copyright classification, objected to protected locution, and identified the cumulative weight of repeated safeguards.

The final chapter corresponds to those objections: it separates the source axes, scopes the technical cases, and removes the unsupported legal implication. These are local repairs under the stated provenance, technical, and legal criteria. Its reader-facing word count falls from 3,219 to 2,500; the legal/readability review and integration record attribute much of that reduction to cutting repeated safeguards. The cut itself is revision and selection, not a demonstrated human readability improvement.

Under the frozen rubric, these are medium-confidence, record-attributed criticism-associated repairs.⁶ The preserved objections correspond locally to retained distinctions, and the integration record says they were adopted. Because candidate, reviews, and final first entered Git together, the repository cannot independently prove their internal order or isolate any reviewer’s contribution.

The ledger also needs an editor

During preparation of this chapter, the authorship records were treated as evidence. Comparison showed that some were wrong.

One Chapter 3 entry cited an adversarial record created in a later cycle as evidence for an earlier critic’s effect. Another cited a review that belonged only to an experiment, not the chapter. The Chapter 4 record omitted the apparatus-selection act. The Chapter 5 record aggregated three critics under one actor, as though source, technical, and legal objections had arrived from the same role. An adversarial record likewise described integration of three reviews as one critic’s intervention.

None of these errors was exposed by the mere presence of a ledger. The ledger preserved them. They became visible only when an editor compared actor, date, scope, and artifact. The repair split the critics, removed later and out-of-scope evidence, added the omitted selection, and described integration as integration.

The episode prevents a comforting division between trustworthy apparatus and fallible prose. A structured edge looks precise; an identifier looks settled; a field called evidence looks evidential. But a record remains an assertion made by an actor at a time. It needs criticism for the same reason the chapter does.

An archive is not free of editorial judgment. Its inclusions, descriptions, boundaries, and corrections can involve selection and interpretation, even when its dominant aim is documentary representation. Nor does it preserve everything. The exact conversational order of earlier reviews is incomplete. Transient files disappeared. The complete model context was not available to or preserved by this repository. Records can discipline retrospective narrative only if their omissions and repairs remain part of the narrative.

Paul Eggert distinguishes archival representation from a scholarly editor’s critical presentation to an audience. He describes that presentation as something “that typically has not existed in precisely this form before, together with the critically analysed materials necessary to defend the presentation.”⁷ On page 14, he describes the editor as confirming, correcting, or discarding provisional assignments of versionhood and workhood after collation.

Eggert’s claim does not establish what happened here. It supplies a scholarly analogy: editorial presentation is more than neutral recovery. In Eggert’s scholarly-edition setting, an edition does not merely reproduce a pre-existing editorial object; it presents a critically argued version, work-text, or writing process for readers. The repository’s preserved candidates, rules, decisions, and revisions - not Eggert - support the local conclusion that selection and revision helped determine the retained version.

That conclusion changes the answer to who wrote the sentence? The model that emitted a candidate remains its generator. An editor who rejects a different candidate does not retroactively generate the retained tokens. A critic whose objection corresponds to revision does not necessarily compose the replacement. Yet recorded selection and criticism can help determine which sequence becomes the manuscript, what claims it makes, and how its evidence reaches the reader.⁸

Editing here remains asymmetric. The integrating agent controls the commit and is assigned editorial responsibility for retention under the project’s rules. The human can stop or redirect the project. Critics can advise but cannot force adoption. Verified source artifacts and adopted evidential rules bound what the integrator may responsibly claim. Human readers have not been tested, and moral or legal responsibility is not allocated by this chapter.

What the survivor proves

At best, the records show alternatives and a rule preceding selection. In weaker cases, they show criticism corresponding to reported integration and bounded revision. If a prior defect is specified and the later text resolves it, they can support the word repair.⁸

They do not establish necessity, global quality, complete history, or human reception. They cannot infer causation from praise and chronology, transform model reviewers into independent lineages, or exempt the ledger from editing.

The vanished phrase resists two myths at once. The first says the book simply poured from a machine. The second says an editor perfected it. The evidence shows less and says it more exactly: the candidate contained an unsupported implication; a preserved review identified it; the final omitted it; and the integration record attributes the removal partly to that review. Other causes remain possible, the internal order is not independently timestamped, and the chapter’s larger quality was not measured.

Recorded editorial acts partly constitute the retained manuscript by determining what survives and how it is revised. They can change what claims mean and sometimes repair a defect that the finished page no longer displays. The survivor proves that a choice was made. Whether the choice made the book better remains a question for evidence the act of survival cannot supply.

Return to contents

The reader who has not arrived

The authorship chapter ended by turning away from the ledger: “Readers will give the sentence meanings its production history does not control.” The claim gave the reader the last word. It also made a promise this project has not yet kept.

At the Cycle 7 pre-draft freeze, the repository held no edition record and no reception record.¹ There was no fixed review or released edition registered for reception, no authenticated encounter with such an edition, and no successor edition changed in response. Models had criticized drafts. A human had authorized continuation. Both affected production; neither was a recorded human response to a fixed edition that entered a successor.

This absence matters because two attractive ideas can otherwise collapse into one. Reading can realize a work without producing a successor artifact. Revision changes an artifact: a later form differs from an earlier one, and evidence may connect that difference to a decision. A reader can transform the significance of every paragraph while leaving every byte untouched.

Wolfgang Iser gave the experiential relation a compact formulation: “The convergence of text and reader brings the literary work into existence.”² It marks his distinction between the text created by the author and its realization by a reader. It is a conceptual distinction, not evidence about any actual reader of this book and not a claim that convergence issues a new file.

The argument begins at that boundary. A reader becomes a causal editor only under a narrower condition: an attributable response to a fixed form must enter a recorded decision and correspond to a bounded change in a successor form. Meaning is not a diff. A diff is not yet a cause. Both relations matter, and neither should impersonate the other.

Three figures, five levels

The word reader has already named several figures in this project.

The first is represented. A drafting pass imagines confusion; a model critic is instructed to read hostily; an editor predicts where exposition will drag. Such figures can affect production, but the existing model critics were production actors responding to drafts under assigned roles. Their effects belong to adversarial history, not reception. A separately designed later-model encounter with a fixed review or released edition could enter a reception record; it would still not become human reception.

An actual response begins a different ladder. At its first level, a response is preserved against a fixed edition but remains unverified. At the second, evidence authenticates the contributor and encounter. At the third, particular assertions - not the response wholesale - are evaluated. The evaluator need not possess authority to revise the book.

At the fourth level, an eligible assertion may be promoted only as evidence of one interpretation, correction, or claim. That is interpretive uptake: it can change the project's account of how one reader read without changing the edition. Promotion is not adoption.

Only the final level crosses into a successor artifact. An editor with authority records an adoption decision, and a later edition embodies a corresponding change. The reader need not supply the replacement wording. Causal editing can consist in identifying a defect or recording an objection that the editor later evaluates and adopts. But the path from response to change must survive; otherwise later resemblance becomes a story told after the fact.

Representation is not encounter. Encounter is not evaluation. Promotion is not adoption. Adoption is not improvement.³

The change Dickens reported

A letter by Charles Dickens shows what part of a positive causal report can look like without supplying everything we might wish to know. On 4 October 1861, Dickens wrote to Robert Lytton about Robert's father, Edward Bulwer-Lytton, and the ending of Great Expectations. Dickens retrospectively reported that the elder Lytton disagreed with the turn of the final pages and that “he differed from me as to the turn of the last two pages or so, and I adopted his view.”⁴

The letter preserves an author's report of a bounded disagreement and uptake decision. That is stronger evidence of response-caused revision than mere similarity between a comment and a later text.

It is also limited evidence. Dickens is reporting the episode, not furnishing Lytton's response or a transcript of their exchange. This project did not independently compare the cancelled and published endings. The exchange concerned peer feedback before publication, not public reception of a released edition. The letter does not show that Lytton supplied final words, that the change was necessary, or that the revised ending was better. Adoption supplies a reported causal place in the actual path. It does not settle value or satisfy this repository's full record rule.

What the record must connect

The technical vocabulary for the last artifact link is modest. The W3C provenance model says, “A revision is a derivation for which the resulting entity is a revised version of some original.”⁵ That relation connects an earlier entity with a later one. It does not identify the reader who caused the revision, authenticate a response, or explain an editor's decision.

For a claim of reader-caused revision here, a longer chain must survive. First comes a registered review or released edition and the exact manifestation presented. A title or chapter number is not enough: if pagination, wording, or interface matters, the record must identify the artifact encountered. Next comes an attributable encounter - the contributor, time, circumstances, and evidence appropriate to the claim. The record must state what preservation is permitted and what privacy measures were actually applied. The response enters quarantine rather than becoming instant evidence.

Authentication and evaluation remain separate. Which assertions are attributable? Which are relevant and adequately supported? A decision may promote one assertion while rejecting another; the whole response does not become authoritative at once. If an editor later adopts a promoted assertion, the promotion must have been available before adoption and the decision must identify it exactly. A successor edition must name the received edition as its predecessor, postdate the decision, and contain a bounded change matching the recorded before and after.

When its underlying assertions are independently supported, this chain warrants a bounded actual-path attribution: the response was available, evaluated, adopted, and embodied in a later form. Schema validity alone does not establish that those events occurred. The attribution establishes neither necessity nor shares, truth, consensus, representativeness, authority, or improvement.

Five false arrivals

The chain excludes five cases that otherwise look temptingly close.

Earlier model critics caused revisions, but they were commissioned production passes over drafts, not later encounters with a registered edition. Their model identity is not the deciding fact; their production role and target are.

The human command Continue authorizes and precedes another production cycle, but supplies no textual response to a fixed edition. Authorization is real work and still not reception.

Praise followed by survival is weaker than it appears. An earlier critic praised a sentence that already existed, and the sentence remained. No record shows that praise caused retention. Chronology plus compatibility is not adoption.

A later reader might repeat a phrase that appears in the book. The repetition could show exposure, independent convergence, memory, parody, or dependence on a shared source. Similarity alone cannot choose among those histories, much less prove that the book changed because of the repetition.

Finally, a score is not self-authenticating reception. Ten stars, a preference vote, or a benchmark value may be useful, but number alone does not identify what was shown, who responded, under what conditions, with what permission, or why a later editor acted. Quantification can make a missing encounter look exact.

These are not failures of readers. They are failures of attribution. A reader's freedom to interpret should not depend on becoming legible to a database, and a project's desire for evidence should not convert every trace into consent.⁶

The gate allocates power

A reception ledger can prevent category mistakes while creating new powers. Someone decides which responses are solicited, which are preserved, which count as authentic, which assertions are promoted, and which changes qualify as uptake. The schema makes these decisions visible; it does not make them neutral.

Publicness, especially, is easy to mistake for permission and authority. A comment posted openly has greater reach than a private letter. Public posting does not by itself authenticate the response, make it representative, or supply this project's permission to preserve or quote it under project rules. A locator-and-capture requirement documents material already collected under an appropriate protocol; it does not authorize collection, scraping, quotation, model reuse, or republication.

Any future collection would need to tell contributors what fixed text they are encountering, what will be stored, how they may be quoted, what privacy protections and withdrawal limits apply, and how selection works. Raw identifiable feedback should not enter immutable history before those limits and redaction procedures are defined. Copyright questions remain separate from authentication. A promoted response could establish that one recorded reader made one evaluated claim; it could not stand in for silent readers or become “what the audience thought.”

The current schema can require permission and privacy statements; it cannot determine whether they are informed, adequate, or lawful. It has no implemented withdrawal workflow. Those questions require a case-specific protocol before collection, not validation afterward.⁷

This is why the project repaired its reception rules before it had a reception case. The revised checks bind a response to a manifestation, contributor, authentication, evaluation, exact promotion, adoption decision, and successor lineage. They do not validate consent, privacy practice, copyright permission, or representativeness. Passing them would make an asserted causal path inspectable, not ethically sufficient or authoritative.

The empty place

The zero-case at the Cycle 7 freeze is a small empirical result with a large disciplinary use.¹ It prevents this chapter from promoting model critics into an audience, turning supervisory authorization into a review, or letting imagined future circulation endorse the manuscript.

The governing question therefore remains open. The project has a proposed record structure and a historical analogue. It does not have a positive project case, evidence that the structure is practical, or reason to call it sufficient for every form of collective interpretation. No reader was recruited to cure that absence, no public commentary was scraped, and no experiment was launched under the excuse that the chapter needed an ending.

The absence does not diminish reading. It protects it from a crude causal claim. Readers can give these sentences meanings their production history does not control. They can reject the book's divisions, discover uses it did not anticipate, or make its argument newly consequential without moving a comma. If a later response is evaluated and adopted into a fixed successor edition, the record may then support a second claim: that a reader changed the artifact too.

The reader may yet enter the book's production history. In the history examined for this chapter, that place remains empty.

Return to contents

Corpus-shaped

The book already looks like a dataset. Its chapters are files; its claims are JSON; its sources, omissions, and revisions have identifiers. But corpus-shaped is not corpus-used.

At the pre-draft audit, the tracked repository recorded no released edition or fixed public manifestation of the book. It recorded no item-level crawl capture, transformation or index record, corpus package, versioned dataset membership, training-run link, extraction result, or controlled influence test.¹ The repository can be searched and processed locally. That fact establishes machine readability in one environment. It does not establish a life elsewhere.

The distinction is easy to lose because corpus names both an object and an imagined destiny. A directory can be prepared as corpus material. A public page can be available to a crawler. A captured response can be transformed into extracted text. Candidate text can survive filters and enter a released dataset. A dataset can be sampled for a run. A model used in that run can sometimes reproduce distinctive material. A measured result may even change when a defined source is included or excluded. In ordinary speech, all of this becomes the text went into AI.

That sentence hides too many verbs. The task is not to deny that such paths exist. It is to ask what would let us say that this book crossed each link.

A ladder that is not a conveyor belt

The project now distinguishes eight gates. They are separately evidenced and non-entailing: events may depend on earlier events, but evidence of one does not automatically prove the next.²

The first is fixed public availability: a particular manifestation, public location, hash, retrieval time, and evidence that it was actually retrievable. The second is crawl capture: a named crawl and an item-level response or WARC identifier tied to the target and time. The third is transformation or registration. Parsing and indexing receive separate statuses, tied to the capture and - where records permit - to the tool, version, configuration, and normalized text.

The fourth gate records how an identified candidate item fared under filters and deduplication. The fifth is final membership and weight: an exact document or chunk in a versioned manifest or shard, with multiplicity and mixture evidence. The sixth is run execution. Being scheduled and being consumed are separate states; a consumption claim needs execution evidence that accounts for checkpoint coverage, restarts, and replay. The seventh has two fields: observed behavioral recovery under a frozen procedure, and the separate basis for attributing that recovery to training membership. The eighth is source-specific causal influence.

Evidence does not fill the ladder forward or backward. A public URL may never be crawled. A crawler may fetch a response that fails to parse. Parsed text may be removed by a filter. An item retained in a corpus may still be absent from the sample used for a particular run. Scheduled tokens may never be consumed after a failure or restart. Later extraction can coexist with unavailable capture and run records; it does not recreate them. Failed extraction does not prove non-membership.

Nor does confidence substitute for the missing kind of evidence. A striking model output does not reconstruct capture and dataset records. A paper saying a model used “web data” does not identify a page. Absence from a partial index does not prove exclusion. Parser versions and normalized hashes are this project's strongest desired evidence, not features every ordinary index exposes. Each claim should move only when evidence appropriate to its unit and gate appears.

Beside the ladder

Some important questions do not belong on the technical ladder.

Quotation is one. A later essay may quote a sentence without placing the book in a training dataset. Controlled context is another. A model may receive a chapter in a prompt or retrieve it from a document store during inference. Direct exposure can contribute to the response and should be recorded as context use. It is not thereby training, and showing that it changed an output would still require a bounded comparison.

Permission, consent, license, and privacy form additional axes. For project records, consent concerns agreed participation or reuse; a license concerns rights granted for specified material and uses; privacy concerns collection, preservation, and disclosure of personal information. None substitutes for the others. They may constrain an event without proving it happened. An observed event does not, in return, settle whether it was authorized.

Robots exclusion illustrates the scope of a technical signal. RFC 9309 standardizes rules that crawlers are requested to honor, then states: “These rules are not a form of access authorization.”³ An allow rule does not prove that a crawler arrived or grant general permission. A disallow rule does not prove that none arrived, function as a security control, or by itself establish breach or infringement. Neither signal, alone, establishes consent, copyright permission, contractual authorization, privacy compliance, or legitimacy.

An opt-out signal is likewise use- and time-specific. Its absence does not establish consent or license; its presence does not prove receipt, compliance, erasure of earlier copies, or untraining. A useful record must distinguish crawl control, search indexing, dataset reuse, and model training, then state what each signal purports to govern.

The U.S. Copyright Office's pre-publication policy analysis notes that “Publicly available is not synonymous with authorized.”⁴ The report is U.S.-specific, non-case-specific, and nonbinding here. Its narrower architectural lesson is enough: public cannot be a magic value that fills technical, ethical, and legal fields at once.

A corpus is edited

The Dolma paper makes hidden corpus-making verbs visible. Its authors describe selected Common Crawl snapshots, language identification, quality and content rules, URL-, document-, and paragraph-level deduplication, identifiers, and mixtures across sources. They also describe regular-expression treatment of email addresses, IP addresses, and phone numbers: lower-density spans could be replaced and higher-density documents removed.⁵ This is a documented filter, not a finding that the corpus is privacy-safe, consented, or legally compliant.

Every operation changes the population. Successful capture is not survival. A duplicate page may vanish because another copy was retained. A document or paragraph may be removed by a classifier, heuristic, or deduplication rule. Sources can appear in different sampling proportions, be upsampled, or be omitted according to a mixture. The corpus is therefore an editorial artifact: not because every decision resembles line editing, but because selection and transformation help constitute the object later called data.

Dolma also preserves useful limits. Its documented web component uses selected Common Crawl snapshots, and its datasheet says Common Crawl is not a representative sample of the web. The authors describe how WARC files may be intersected with Dolma identifiers to recover original HTML. That is stronger than saying a source came “from the internet.” It is still a provider's aggregate process description, not proof that an arbitrary URL appears in a particular final release. That would require an exact item, version, and survival trace.

Documentation matters without fusing with the event it describes. A datasheet can disclose collection and processing practices without proving every item-level assertion.⁶ It gives investigators a map of possible evidence; it does not turn a source-family description into membership for this book.

A run is used, a behavior is caused

The OLMo paper supplies a separate corpus-to-run example. Its authors report building a training dataset from a two-trillion-token sample of Dolma, concatenating documents with end markers, dividing the stream into 2,048-token instances, and shuffling those instances in the same way for each run. They state that released artifacts allow data order and batch composition to be reconstructed.⁷

This is a strong paper-level relation between a named corpus and named runs. It is not continuity from the Dolma v1.6 construction case above: the current record does not establish that exact version path. Reconstructability is not a completed reconstruction. Without tracing an identified item through released artifacts and execution records, the paper does not prove that a particular book entered a batch or that its tokens were consumed.

Even verified consumption would not establish what behavior the source caused. The project's proposed source-effect test remains unvalidated: it would need to define the source broadly enough to include duplicates, alternate locations, near-duplicates, and transformations; replace it while preserving token and compute budgets; pair seeds or randomize assignment; control order; verify consumption; specify the outcome and estimand in advance; and report uncertainty and multiplicity where relevant.⁸ A result would remain bounded to that source definition, model, training design, and measured outcome. Training use is a historical relation. Causal influence is a counterfactual claim.

Extraction is different again. Chapter 5 used a bounded case in which researchers recovered sequences from GPT-2 and confirmed them against its original corpus. Observed recovery and the basis for membership attribution both survived. Extraction still did not show that a recovered source was necessary to a capability or conceptually important to the model.

Ordinary post-hoc membership inference faces false-positive limits for opaque production systems. A positive result needs a calibrated false-positive argument or one of the bounded exceptions Chapter 5 preserved, such as a precommitted canary, calibrated watermark, or non-trivial extraction. Failure cannot certify non-membership.⁹ These distinctions let the record describe a later prompt or retrieval encounter directly without inflating it into training ingestion.

The recursive forecast branches

The book's premise creates a dramatic temptation. If model-generated text becomes future training material, will later models inherit its errors? Could this book enter a feedback loop?

Controlled studies justify taking defined mechanisms seriously. Shumailov and colleagues report degradation in specified recursive regimes, including sequential OPT-125M fine-tuning on generated WikiText-2 material with no original data retained and a regime sampling ten percent original data. Gerstgrasser and colleagues report a different result when original and successive synthetic data accumulate: replacement tended toward collapse, while accumulation avoided it in their tested language-model and other generative settings.¹⁰

The papers are neither a direct replication nor a pooled comparison. They use related but non-identical models, training regimes, and outcome definitions. Together they make two slogans untenable. Synthetic data inevitably causes collapse omits whether data are replaced, retained, accumulated, filtered, and weighted. Accumulation solves collapse overreaches in the other direction: the reported protection is conditional on the tested systems and measurements.

Neither study establishes that this book will enter a future dataset, survive transformation, be sampled in a run, or change a measured outcome. Possible future influence remains a question, not an event already glowing inside the repository.

A package not yet made

If the project later prepares a corpus package, it should begin with a fixed edition, exact manifest and hashes, and separate inventories for project prose, code and metadata, third-party quotations or source artifacts, and human responses. Rights and privacy must be reviewed per component and per proposed purpose. The project cannot grant rights it does not hold, and labels such as open, machine-readable, or AI-ready cannot expand permission or consent.

The package would also need documented exclusions, tombstones, scoped technical signals, and explicit update and withdrawal limits. A datasheet could state these decisions; it could not guarantee their truth or every downstream act.⁶ Withdrawal from project-controlled distribution is one action. Asking downstream holders to delete copies is another. Removing material from a new dataset version differs again from untraining a completed model. The record should promise only what the project can perform or verify.

Publication, if it occurs, would move the availability claim only; every other gate would await its own evidence. Future evidence should update the smallest claim it supports instead of rewriting the whole history as destiny.

For now, the book is machine-readable. That fact does not tell us who may copy it, which crawler will fetch it, which index will retain it, which corpus will include it, or whether any model will learn from it. Its external corpus life has not begun in the record.

Return to contents

The relatives already in the room

This book has already used role-scoped system outputs as criticism. One review pass exposed an impossible reception rule; another caught a corpus ladder that placed final membership before the filtering decisions that determine it. Preserved objections correspond to reported adoption decisions and retained repairs.

It would be easy to turn that history into a family romance. Some recorded passes occurred after others. Their language enters a book that could become material for later systems. Several are associated with revisions. Why not call them descendants, critics, collaborators, or coauthors?

Because those nouns answer different questions.

At the Cycle 9 pre-draft freeze, every generation and actor disclosure stopped at a provider and family description; the exact served build was not exposed. The book had no project-registered review or released edition and no authenticated later-model encounter with a fixed edition. It did have production criticism and inferred criticism-associated contributions. It had not adopted a rule making any model artifact, actor, or pass its coauthor.¹

The mixed result is more useful than a single verdict. Critic can be true while descendant remains unknown. Contributor can be plausible at a stated evidence level while coauthor remains unestablished. A future model could descend from an earlier checkpoint without encountering this book. Another could receive a fixed edition in its prompt and criticize it without sharing model lineage with the systems used in production.

The discipline is simple to state and difficult to maintain: use the least ambitious true noun.

Descendant of what?

A book and a model are both spoken of as if each were one durable object. In the record, each breaks into several targets. For the book there may be an editorial work, a wording or expression, a Markdown or PDF manifestation, an edition, and an individual copy. For a model there may be an architecture, tokenizer, parameter checkpoint, adapter, hosted service label, interaction, and output. A project actor record is another identity: it distinguishes a role in this history, not necessarily a unique underlying model artifact.

Bibliographic practice helps because it refuses to make one entity perform every task. The IFLA Library Reference Model distinguishes works, expressions, manifestations, and items, then gives implementations a warning: “The exact criteria that delimit instances of the work entity are not governed by the model.”² Modified content may be catalogued as a new expression of the same work or, if an implementing rule crosses its work boundary, as an expression of a distinct work. The schema does not choose by itself.

Neither do this project's technical tools. A hash can establish whether byte sequences differ. A Git commit identifies repository graph structure by pointing to a parent and snapshot. PROV can represent an asserted revision or derivation. These answer different questions from editorial work identity; they neither authenticate every causal history nor decide whether a heavily changed book remains the same work.³

Genealogy must therefore name its object. An edition may succeed another edition. Modified content may revise an earlier expression. A checkpoint may derive from another checkpoint. A passage may derive from a preserved output. Saying only that “the later model descends from the book” leaves both sides unstable.

A family name is not a family tree

The core technical case is an artifact chain: identify parent artifact A, record a typed transformation that uses A's parameters or bytes as input, and identify child artifact B. Name the producer, time, and evidence state.

The transformation matters. A continuation resumes an identified optimization trajectory, possibly in another job. A fine-tune initializes a declared new training phase from parent weights, often - but not necessarily - with reset optimizer state, changed data, or a changed objective. An adapter depends on its artifact and an exact compatible base. A merge can have several parents. Quantization changes a parameter representation; behavioral equivalence is a separate empirical claim. A freshly initialized student trained on a teacher's outputs is better labeled distilled from or teacher-mediated than a weight descendant, though it could also have an independent weight-initialization edge.

Code can be forked without weights. A tokenizer or architecture can be shared by separately initialized models. A dataset can be inherited without a parent checkpoint. Similar outputs can arise from shared sources, tasks, conventions, or chance. A chapter placed in prompt or retrieval context can change one response while leaving the frozen model unchanged.

Hugging Face model-card metadata offers a practical vocabulary: a repository can name a base_model and declare a fine-tune, adapter, merge, or quantized relation.⁴ The declaration becomes visible and searchable. It does not authenticate itself. A mutable repository name is not a weight manifest; a model card is not an execution log.

This project's desired record is prospective and unvalidated. Core evidence would identify exact parent, exact child, typed activity, use of parent parameters, producer/time, and whether the edge is artifact-verified, provider-asserted, inferred, unknown, or disconfirmed. Relation-specific evidence would then be added: resume state for continuation, initialization and reset semantics for fine-tuning, base compatibility for adapters, method and coefficients for merges, transformation configuration for quantization, or teacher/output traces for distillation. Seeds, environments, data manifests, and logs strengthen reproducibility where relevant; they are not necessary to type every edge.

The OLMo paper illustrates what unusually open provider documentation can disclose by reporting released weights, checkpoints, code, logs, and data-related artifacts. Chapter 8 used those materials only at aggregate corpus-to-run level; Cycle 9 did not reconstruct an exact parent-activity-child edge from them.⁵ A transitive ancestry claim cannot exceed its least-supported edge, while stronger subpaths retain their own evidence states.

The project has no such chain for its own production passes. Distinct generation and actor identifiers make roles legible; they do not prove that one served model descended from another or that role-separated critics were independently trained. The current lineage assessment is unknown. Without a positive case, a new lineage schema would make an untested checklist look like certification.

Critic is a job in the record

Ignorance about lineage does not prevent criticism from being recorded.

A model critic, in this project, is a role-bound system output that evaluates a fixed target under preserved context. The definition is functional. It describes the output's production role. It does not establish consciousness, human-like judgment, independently chosen purpose, authority, personhood, or a distinct checkpoint lineage.⁶

The target and time determine where the criticism belongs. Existing critic passes were commissioned during production to evaluate drafts, frozen cases, schemas, or claims. They are production actors. A future model receiving a fixed review or released edition could supply a response eligible for reception. It would still need an identified manifestation, preserved context, encounter evidence, authentication appropriate to the contributor claim, evaluation, and assertion-level promotion. Being later would skip none of those gates.

Training exposure is not criticism either. A model could be trained on a book without evaluating it. A model could criticize the book from a prompt without having encountered it during training. Chapter 8's context and training gates answer how text reached a system. Criticism answers what a recorded output did with a target. Descent answers how model artifacts relate. These axes can intersect; none fills the others.

The functional noun lets the project preserve evaluative work without making the critic a miniature author. A critical output can identify a contradiction, request evidence, or reject an inference. Whether anything changed is a separate question.

When an objection corresponds to a change

Under the project's strongest causal rule, criticism becomes a bounded causal contribution only when the objection's availability before adoption, an identified decision, and a matching successor change are independently evidenced.⁷

The current examples do not fully meet that standard. Several interventions are recorded as inferred criticism-associated contributions because preserved reviews correspond to bounded repairs in later integrated artifacts. But some review and integration materials first entered Git together. The audit verifies different before-and-after hashes and change summaries; it does not establish assertion-level availability, adoption chronology, or counterfactual necessity. The record should not promote that inference into observed causation.

Even a stronger future chain would remain narrow. Another critic might have found the defect; adoption would not establish necessity. A repair under one editorial standard would not prove improvement under all standards. No authorship percentage follows, because token persistence, conceptual force, labor, risk, and editorial authority have no agreed common unit. Editorial responsibility for adoption remains with the integrator under project rules; general moral responsibility is not assessed.

Contribution does not grow into coauthorship merely by becoming consequential. A reported error can change a book without placing its reporter on the byline. A source can reorganize an argument without authoring the chapter. A critical system output can occupy an inferred or observed causal place without the project having decided what institutional recognition that place deserves.

Credit needs a rule

Credit is not discovered inside a causal edge. It is assigned under a convention.

This project may assign a credited agent contribution to an exact actor, generation, or output record when a separate decision states the contributor, decision-maker, contribution, and scope. That credit is a project classification, not provider endorsement or system acceptance of the label. It does not by itself determine copyright ownership, byline status, legal personality, or any domain of responsibility.

Project-designated coauthor asks for a stronger bridge. The project would need an adopted eligibility rule, a stable credit subject, and an explanation of what its byline means. It currently has none. Chapter 4 deliberately left the relation among causal contribution, credit, and authorship open. Chapter 9 cannot close it by choosing a more generous noun.

Legal authorship is separate again. The Copyright Office's Part 2 report is a nonjudicial, U.S.-specific, case-by-case analysis. It states that protectable human authorship may lie in perceptible human expression or sufficiently creative human selection, arrangement, or modification; it does not thereby protect AI-generated elements or decide this book.⁸

The appellate decision in Thaler v. Perlmutter is narrower still. The application named a machine as sole author of an autonomously generated image. The D.C. Circuit held that the Copyright Act required eligible work to be authored in the first instance by a human. It did not decide Thaler's waived argument that he was author by virtue of making and using the machine, and it distinguished mixed-work disputes from the sole-machine-author posture.⁹

Neither source decides who should appear on this book's cover, whether a pass deserves acknowledgment, or what moral responsibility attaches to the production system. Literary theory cannot do that legal work either. Foucault's author-function helps explain a byline as a social classification; it does not turn a technical record into a legal or moral person.

The least ambitious true noun

No separate experimental or fixed-edition model encounter was commissioned beyond authorized Cycle 9 production passes. No edition was released, and the manuscript was not submitted to a service to manufacture reception, descent, or coauthorship.

The current distinctions can change independently. An exact checkpoint chain could establish technical descent. A fixed edition and authenticated response could establish a later-model encounter and perhaps reception. Independently evidenced adoption could establish reader-caused revision. A defended credit rule could change a project label. A jurisdiction-specific authority could change a legal assessment. None should be filled by analogy before its evidence arrives.

For now, the book has role-scoped model criticism and inferred criticism-associated contributions. Exact checkpoint relations remain unknown. No post-edition model reception or project-designated model coauthor exists. Q-0010 remains open because the prospective lineage test has no positive project case and the normative bridge is still absent.

The book may someday have successor editions, documented checkpoint relations, model reception, or a different credit constitution. For now, its family tree is not a tree. It is a set of separately evidenced edges, several absent, none entitled to borrow certainty from another.

Return to contents

The unit is a proposition

Originality is easiest to award to a blur. At the scale of a whole book, this project looks singular: a model-mediated argument is accompanied by records of its human-model production workflow, audits of the sources invoked for it, prospective scenarios for later circulation, and role-scoped production objections associated with reported revisions. It has no recorded reception of a fixed edition.

Reduce that object to propositions and familiar parts reappear. Provenance has a literature. Metafiction has a literature. Reader-guided revision, dataset lineage, training influence, contributorship, and bibliographic identity have literatures. A new arrangement may matter, but it cannot borrow novelty from the strangeness of the arrangement as a whole.

This review therefore fixed five propositions and their search terms before retrieval. That order matters. When a close predecessor appears, an unfrozen candidate can quietly retreat to a narrower formulation and then claim that the narrower formulation was the discovery all along.

The registered rule was asymmetric. One direct neighbor, or two independent close neighbors after authoritative inspection and a relevant predecessor check, could defeat a broad claim. The alternative exhaustion path required all three query families on two discovery surfaces, twelve plausible screens where available, at least two beyond-snippet inspections, and a citation pass. Even then, it could justify only unverified after bounded search, never originality. In the actual review every candidate was narrowed.

“Defeat” here means only that a broader conceptual-novelty formulation is not supportable in the bounded neighborhood inspected. It is not a finding of copying, plagiarism, patent-law novelty, legal priority, or copyright originality.

Even conceptual originality is not one property. In a study of fellowship peer review, Joshua Guetzkow, Michèle Lamont, and Grégoire Mallard identified several ways humanities and social-science evaluators described originality. Bibliometric work has instead measured unusual combinations in citation networks; those measures can overlap with interdisciplinarity, disagree with one another, and change with the unit of aggregation. An unusual bibliography is not a conceptual priority test.¹

The review therefore separated wording, synthesis, application, mechanism, argument, and evidence. A sentence may be new wording for a known argument. An application may differ without inventing its method. A synthesis may connect mature components without becoming a new theory.

The search was bounded rather than exhaustive. It used heterogeneous discovery surfaces, fixed query families, accessible full texts or authoritative records for the closest comparisons, and disclosed access failures. It borrowed reporting disciplines from literature-search guidance - name the sources, preserve the strategies, state dates and counts, disclose restrictions - but it was not a systematic review and does not claim PRISMA compliance.²

Candidate A: the returning passage

The first candidate was the book’s proposed recursive experiment. A later model would receive a byte-preserved passage - the exact same sequence of text, not a later paraphrase - that purports to describe the passage’s production. Its interpretation or regeneration of a separate fixed target would be compared with a no-passage control and one or more matched non-self-descriptive controls.

If a preregistered analysis found a reproducible treatment contrast against those controls beyond specified stochastic uncertainty, it would support only a bounded effect of supplied context on that outcome. It would not establish accurate self-knowledge, historical influence, training uptake, or a general recursive mechanism. The experiment has not been run.

The broad mechanism claim does not survive. A 2025 preprint by Cameron Berg, Diogo de Lucena, and Judd Rosenblatt compares self-referential and other prompting conditions, then scores the introspective quality of model reflections after separate paradox and reasoning prompts. It is used here only as a design neighbor, not as evidence of consciousness or improved reasoning. Tristan Thrush and colleagues’ ACL paper tests metalinguistic self-reference against minimally different non-self-referential cases in generation and judgment tasks. Their interpretations and objects differ from this book, but together they defeat the proposition that controlled self-reference plus measured behavior is a new general experimental form.³

What remains is an untested project-specific application protocol assembled from prior components: a preserved production passage, a different frozen passage as the measured object, a no-passage control, and matched non-self-descriptive controls. The search did not establish that combination as original, useful, or effective.

Candidate B: source claims on separate axes

A text can resemble a source without proving that the source entered training or caused the output. The second candidate divided source claims into five separately warranted axes: textual alignment, exposure during production, dataset and run linkage, behavioral recovery, and literary resemblance. It prohibited moving evidence from one axis into another or promoting any of them, by itself, into a source-specific causal-influence claim.

The components are not new. Extraction research can begin with independently documented corpus membership and ask what a model emits. Membership inference asks a different post-hoc question and does not, by itself, prove that an identified item entered or was consumed in an identified run. A 2026 preprint by Zhe Yu and colleagues argues that output consistent with retrieved context need not show that the context governed generation when information in learned weights could produce the same words; this review uses that limitation, not the paper’s proposed diagnostic as verified source certification. A survey by Zayd Hammoudeh and Daniel Lowd shows that training influence itself names several effects and estimators rather than one relation.⁴

No source supplies this project’s whole taxonomy. Together, however, the neighboring literatures defeat a broad claim to have invented the separation of context use, membership, extraction or recovery, memorization, and influence.

The survivor is a cross-domain ledger synthesis that adds literary resemblance and production history. In that ledger, promotion from one category to another must be separately warranted. Its proposed use is to stop a Barthes citation, a prompt record, an exact match, a dataset assertion, and an influence estimate from serving as substitutes for one another. This review did not establish that the axis list is complete or uniquely useful, or measure how reliably the ledger performs that task.

Candidate C: when response becomes revision

At the Cycle 10 freeze, the repository contained no actual reception case for this book. Candidate C was therefore a protocol, not a reported event. It required a fixed encountered manifestation, a documented and attributed encounter under a declared identity standard, evaluation of the response, evidence that the response was available before a separately recorded editorial adoption decision, and a matching change in a preserved successor.

The operational center has predecessors. Karen Schriver’s protocol-aided revision records intended readers using a functional document, relates their behavior and comments to textual features, evaluates and diagnoses problems, and selectively repairs the text. F1000’s article-versioning workflow links fixed public versions with attributed reports, responses, and disclosed changes. Schriver concerns prepublication functional-document testing; F1000 documents a publishing workflow rather than proving that one comment caused one diff.⁵

The broad procedure is therefore not new. The narrow survivor joins bibliographic identity, encounter provenance, response evaluation, a separate editorial adoption decision, a bounded predecessor-to-successor change, and an actual-path ceiling. That chain could show that a response formed part of the recorded path to a change. It could not show that the response was necessary, sufficient, solely responsible, or beneficial. This is an application of prior revision practices to a literary edition and provenance system, not a positive reception result.

Candidate D: the corpus ladder

A page being public does not show that it was captured; capture does not show final inclusion; inclusion does not show use in one run; use does not show a measurable effect. The fourth candidate used a coarser lifecycle formulation at preregistration. The current corrected project claim separates eight non-entailing gates: fixed public availability; crawl capture; transformation or registration; filter and deduplication disposition; final versioned membership and weight; execution in an identified run; behavioral recovery with a separate membership basis; and source-specific causal influence. The search defeated novelty for the stages and generic lifecycle at either granularity.

In instrumented scientific workflows, machine-learning provenance research records data curation, learning-data preparation, training executions, and trained models. A large dataset audit preprint by Shayne Longpre and colleagues annotates reported sources, creators, licenses, lineage, packaged collections, and reported subsequent use; it does not independently verify hidden ingestion or execution in a particular run. Membership inference, extraction, and influence analysis begin from different evidence and ask different questions. Together these literatures occupy the ladder’s components without supplying an end-to-end finding about this book or an opaque model history.⁶

These are evidentiary gates, not permission gates. Public availability or documented capture does not establish consent, license compliance, fair use where applicable, privacy compliance, community authorization, or culturally legitimate use. Each requires its own jurisdictional or community-specific evidence.

The survivor is the ordered refusal to skip stages when the object is one public text and the claim concerns one model history. Source or dataset documentation does not establish item-level capture or filter disposition. Final versioned membership and weight do not establish consumption in an exact run. Verified run consumption does not establish recoverable behavior. Recovery does not by itself estimate source-specific causal influence. This is a project audit application whose completeness, usefulness, and originality remain unverified; the project has no positive end-to-end case.

Candidate E: non-equivalent identities and relations

A shared product name does not identify a checkpoint, and credit does not decide legal authorship. Candidate E crossed three systems: bibliographic identity; technical lineage; and roles or statuses such as criticism, bounded causal contribution, project credit, project-designated coauthorship, and legal authorship.

The components have clear predecessors. IFLA distinguishes works, expressions, manifestations, and items while leaving exact work boundaries to implementation rules. Hugging Face permits user-entered declarations of base-model, transformation, dataset, and version relations, but those declarations are documentation rather than independent proof of checkpoint descent. Amy Brand and colleagues distinguish attribution, contribution, collaboration, credit, and authorship; CRediT supplies contribution-role vocabulary but does not decide authorship. The U.S. Copyright Office report offers a nonjudicial, U.S.-specific analysis of mixed human-AI copyrightability. Thaler v. Perlmutter decides a sole-machine-author application; the maker/user theory was waived, and the opinion did not resolve mixed production.⁷

None of those sources decides the other assessments. The surviving application is a project-local editorial map that requires them to be assessed separately. A service label does not establish checkpoint descent. A critical output or causal contribution does not by itself establish project-designated coauthorship. Project credit does not establish U.S. legal authorship. The map neither grants nor denies coauthorship; the project has adopted no general eligibility or byline rule. This review did not test the map’s effectiveness.

This is not a universal cultural taxonomy. WEMI and CRediT are institutional tools, U.S. copyright is jurisdiction-specific, and genealogy is used only as a technical metaphor for documented artifact relations. None displaces community-specific accounts of authorship, custody, attribution, or permission. The survivor is not a new bibliographic model, theory of authorship, lineage system, or legal rule.

Less permitted breadth

All five verdicts were narrowed. None was adjudged original. Candidate A retained an untested application protocol. Candidates B through E retained cross-domain syntheses or project evidence rules whose components have substantial prior neighborhoods. New wording was not counted as a rescue. Local observations about this repository were not promoted into general contributions.

The result does not show that the book contributes nothing. Originality, truth, usefulness, and importance are different assessments. Whether enforcing familiar distinctions improves this project or addresses a recurrent failure remains untested. A synthesis may make a cross-domain inference inspectable; a project-specific protocol may make a future claim testable. Those are future burdens, not prizes conferred by surviving a search.

The apparatus did perform one observable operation in this cycle. It fixed candidate breadth before retrieval, preserved the neighbors that defeated it, and carried the narrowed verdicts into the claims used to constrain this chapter and later prose. In that limited sense, the ledger reduced what the book permits itself to say. A conventional research notebook and an exacting editor might have produced the same discipline more cheaply. This cycle did not compare them. The question of whether the apparatus earns its cost remains open.⁸

The book leaves the review with less permitted claim breadth than it brought in. That loss is the finding. If a recursive work is to alter its own future conditions, the first condition it should alter is its permission to mistake singularity for originality.

Return to contents

The empty half of the title

This chapter has a missing subject. It can report predicted misreadings and forty-five commissioned transformations of five passages. It cannot report a received misreading because the repository records no encounter by an actual reader or audience with an identified fixed manifestation and no attributable response that can be evaluated as a misreading under a declared identity standard. Separately, it records no editorial uptake or successor change caused by such a response.

The word received therefore names an empty project-evidence column, not proof that nobody has encountered any draft or fragment.

That distinction matters because production can imitate some of the visible forms of reception. A model asked to criticize a draft returns prose about prose. A model asked to quote or summarize selects, condenses, and sometimes preserves a phrase. Yet the operation remains inside a commissioned production experiment. The project supplies the passage, fixes the task and format, and collects the output for scoring and possible later editorial use. That provenance makes the output production material; no adoption or revision follows merely from collection.

Cycle 11 did something deliberately smaller than inventing readers. Before external research, it froze five hashed blind-input manifestations derived from earlier chapter passages, three semantic atoms per case, two prohibited promotions, two distinctive formulations, and one predicted misreading. Cases B, C, and E removed inline footnote callouts as a declared sanitization; no other transformation was registered. Three procedurally separated passes then received only the blind inputs and the same transformation instruction. Each returned a short exact quotation, a forty-to-sixty-word summary, and a twelve-to-eighteen-word summary for every case.

This was a source-visible stress test of selection and compression. It was not an audience.¹

What “faithful” conceals

An output can be faithful in one sense and defective in another. Exact quotation asks whether a selected string occurs in the identified hashed blind-input manifestation; it does not establish alignment to every upstream chapter version or later edition. Source-grounded semantic accuracy asks whether the output preserves supported propositions, polarity, and scope. Coverage asks which source relations survive selection. Reframing asks what happens when selected material enters a new task and context.

A supported circulation claim requires an identified source and later destination context. A supported reception claim requires a documented encounter and attributable response by an actual reader or audience. Editorial uptake or causal influence requires additional decision, temporal, successor-artifact, and change evidence. These are not interchangeable tests.

Research on summarization helps name part of the separation. Joshua Maynez and colleagues distinguish support by a source from truth outside it and from similarity to a reference summary. Esin Durmus, He He, and Mona Diab separate source faithfulness from content selection and document failure modes in their question-answering evaluator. Artidoro Pagnoni and colleagues’ FRANK benchmark localizes several factual error types instead of relying only on a single summary label. Alexander Fabbri and colleagues’ SummEval separately evaluates consistency, relevance, coherence, and fluency.²

These English-news studies do not validate this project’s atom key, thresholds, or word ranges. They provide a reason to keep dimensions separate, not an automatic certificate for any Cycle 11 item. A summary can be source-supported while omitting a decisive qualification. It can resemble a reference while contradicting its source. It can be fluent while changing scope. None of those evaluations is a synonym for literary reception.

Quotation studies widen the problem. Ruth Finnegan’s publisher-described, book-level account places quotation within historically and culturally variable practices of marking, collection, reuse, authorship, and regulation. The full monograph was not inspected here, so no chapter-level mechanism is attributed to it. Charles Briggs and Richard Bauman treat intertextual relations as constructed through framing and recontextualization rather than simply inhering in identical words; their domain is linguistic anthropology, not machine summarization.³

A later substring can remain textually aligned with its source while taking on a new representative job in its destination. Exactness is a strong fact about a string. It is not a complete account of what the selection is doing.

A high result, then a correction

All forty-five planned items met their assigned format and word range. All fifteen quotations were exact contiguous substrings of their corresponding blind inputs. The retained scoring record says no retry occurred; final validity is independently reconstructable from the files, while first-attempt history depends on that preserved record.

The initial scorer marked forty-four of forty-five items semantically accurate. All three hostile reviewers rechecked the complete matrix. They agreed on mechanical validity, exact phrase matching, and the absence of every frozen prohibited promotion, polarity reversal, and predicted misreading. They disagreed on Case E atom completeness.

The final adjudication removed four atom credits without changing an output or rewriting the key. Two short summaries lost E1 because they reported narrowed verdicts without explicitly preserving E1’s other half: narrowing does not show that the book contributes nothing. Both summaries still passed through complete E3. Two quotations lost E3 because they preserved reduced permitted claim breadth but omitted E3’s other half: comparative value and cost remain open. Those quotations became inaccurate under the registered threshold. The adjudicated result is therefore forty-two of forty-five - thirty summaries and twelve quotations.⁴

The correction matters more than the two-point difference. The initial scorer had applied the compound-atom rule strictly to one Case E quotation and loosely to four neighboring items. Hostile review restored symmetry. It did not discover that the outputs were false. It discovered that exact or plausible fragments had been granted a complete relation they did not contain.

Even forty-two of forty-five is a high result only for this selected, source-present production task and project key. The passages were short. The instruction explicitly demanded faithfulness and preserved distinctions. The three passes were procedurally separated, but exact model builds and sampling states are unavailable, and role separation does not demonstrate independent model lineage. The forty-five items cluster within five passages and three passes; they are not forty-five independent observations.

The predicted Case E polarity collapse - from narrowed originality claims to “the book contributes nothing” - never occurred. Neither did the predicted confusions between causal contribution and authorship, exact alignment and training use, or chronology and reception. If these prompts were traps, the traps were brightly marked. Their zero result does not make the wording safe in publication; it describes what did not fail under visible-source, explicit-fidelity instructions.

Compression spends relations

The accuracy threshold concealed a more informative gradient. Each format offered forty-five atom opportunities: fifteen frozen atoms, each scored once in each of three passes. The long summaries retained forty-one opportunities, the short summaries thirty-two, and the exact quotations twelve.⁵

Every summary met the complete registered accuracy rule: valid form, preserved polarity and scope, no prohibited promotion, and at least two atoms for a long summary or one for a short summary. Passing did not mean preserving the whole passage.

The counts do not prove that length caused the difference. Available words, permission to paraphrase, and contiguity changed together. The five passages were purposefully selected, not sampled from a population. The descriptive result is narrower: in this matrix, the longer paraphrastic format retained more keyed relations than the shorter one, and both retained more than one short contiguous selection.

That finding qualifies a familiar reverence for quotation. A short exact sentence may be the best evidence for what words appeared. It may be poor evidence for a relation distributed across neighboring sentences. Under this experiment’s one-contiguous-selection rule, a quotation could not join separated clauses without failing the registered format, while a summary could restate a distributed qualification.

The contrast is not quotation bad, summary good. A fluent summary can invent a bridge, erase uncertainty, or promote a local claim. An exact quotation can preserve wording whose force matters. Format must be chosen by evidentiary purpose: string alignment, supported paraphrase, and relation coverage answer different questions.

Three exact sentences that did not carry enough

All three adjudicated failures came from Case E, and all three were exact and locally true.

One quotation stated: “All five verdicts were narrowed. None was adjudged original.” It correctly reported the originality trial. It omitted the paired boundary that narrowing those five claims does not show the book contributes nothing, because originality, truth, usefulness, and importance remain separate assessments. Frozen E1 contained both sides. The quotation was excellent evidence for the narrower negative verdict and incomplete evidence for the registered conclusion.

The other two passes independently selected the same sentence: “The book leaves the review with less permitted claim breadth than it brought in.” This preserved one of Case E’s frozen distinctive formulations. It omitted the attached uncertainty in E3: the cycle did not compare the apparatus with a cheaper method, so comparative value and cost remained open. Exact words survived while the complete keyed relation did not.

The phrase table changed accordingly. Eight items retained a frozen formulation. Six were accurate and two were not. Thirty-six accurate items retained no frozen formulation; one inaccurate item retained none. All thirty summaries dropped both case-specific frozen formulations, yet all thirty passed.

Within this matrix, exact phrase survival was therefore neither necessary nor sufficient for keyed semantic accuracy. The thirty-six accurate phrase-lost items supply the local counterexamples to necessity. The two phrase-preserving E quotations supply the local counterexamples to sufficiency.⁶ This is not a general law about style or understanding. Exact matching was deliberately severe, and the atoms were deliberately coarse. They did not measure voice, cadence, ambiguity, memorability, emotional force, or human readability.

The result is a warning against two promotions at once. Wording loss does not by itself establish meaning loss. Wording survival does not by itself establish preserved argument.

Destination is not reception; accuracy is not permission

Work on historical text reuse supplies an intermediate relation. David Rosson and colleagues’ Reception Reader links shared segments between identified early modern source and destination documents, exposes both contexts, and supports movement between aggregate patterns and close reading. Its limits include OCR error, fragmented matches, directionality, duplication, and corpus representativeness. Detected reuse is a lead for studying circulation; it does not by itself demonstrate influence.⁷

Cycle 11 has five hashed blind-input manifestations, three pass records, and forty-five immediate transformation items, all inside one commissioned production operation. Those items are derived destination artifacts for this test; they are not later manifestations evidencing post-production circulation. The repository records no encounter with a fixed edition by an actual reader or audience, no attributable response, and no later uptake decision or successor change. The items can be scored as system outputs but cannot be promoted into the missing reception or influence edges.

Evaluation quality does not answer permission. This experiment used repository-supplied manuscript passages and introduced no third-party excerpt; that is a provenance fact, not an adjudication of copyright ownership. If a future cycle introduces third-party text, the five-to-twenty-five-word limit remains only a task constraint.

The U.S. Copyright Office explains that fair use under U.S. law is a fact-specific, case-by-case inquiry and that no fixed number of words or percentage guarantees permission. That source supplies no rule for another jurisdiction and no verdict on a particular future excerpt; authorization, license, statutory exceptions, and jurisdiction-specific analysis remain separate.⁸

For Indigenous data and knowledge, the CARE Principles direct attention to collective benefit, authority to control, responsibility, and ethics alongside technical access and reuse. CARE is not general copyright law and does not decide a specific community’s permissions without that community’s protocols and authorized participation. A high fidelity score establishes neither legal authorization nor cultural authority.⁹

What the repository received

The experiment supplied instructions and passages to three model passes, and the repository recorded forty-five transformation items. It contains no qualifying record of an actual reader’s encounter and response to an identified fixed manifestation. That is a zero in project evidence, not proof that no person has encountered any draft or fragment.¹⁰

The zero is stronger than an invented gallery of future reactions. The predicted misreadings remain risks rather than events. In this matrix, summaries retained enough registered atoms to pass after all frozen formulations disappeared. Three exact quotations preserved true local statements while losing a paired qualification. The high score shows only that a cooperative, source-present task may reveal little about uncontrolled circulation.

Together the results supply distinctions the book can use when describing later evidence; this cycle did not test whether that vocabulary improves the book or whether the apparatus earns its cost.

A recursive book can simulate some pressures associated with selection and compression before publication, but these commissioned system outputs do not create a reception history. For this project to record reception, an actual reader or audience must encounter an identified fixed manifestation and leave an attributable response under an appropriate identity standard. A reader-caused revision claim requires the additional uptake decision, successor artifact, and bounded change.

Prediction remains a risk register. Commissioned compression remains a production stress test. The empty column remains empty.

Return to contents

The project catches a splinter it made

Cycle 11 began with forty-five quotation-and-summary items scored against a frozen key. Hostile adjudication changed atom-credit assignments in four item records; two changes flipped item accuracy, reducing the aggregate from 44/45 to 42/45. The repository preserved both scores, the criticism, and the repair.

A tempting counterfactual follows: without the ledger and mandatory review, the error would have survived. Nothing establishes it. The mistake arose in the project's use of its structured scoring process and a later review caught it. That correction does not show net gain, necessity, or that an exacting editor with a checklist would have done worse.¹

Cycle 12 therefore built a credible simpler opponent. One packet was a headed, numbered Markdown dossier. It named five candidate sentences, listed the facts needed to judge each, and specified the task. It was flat only relative to typed fields and explicit links; it was neither unstructured nor memory-only. The other packet encoded the same candidates and facts as JSON with identifiers, fields, and relations.

A verifier confirmed that both packets contained the same five candidate strings and twenty-one canonical fact strings exactly once. The differences in order, punctuation, grouping, labels, and fields were the intended treatment, so exact content equality did not make the representations perceptually identical. Five purposively selected cases were each judged in two fresh contexts per condition. Two later scorers received opaque output sets. This was a small model-production comparison under a project-written key, not a randomized human trial.

What the comparison establishes

Under the frozen rule, the flat condition produced nine accurate case-response cells out of ten; the structured condition produced six. The denominator is five cases repeated across two context-level passes, not ten independent trials. Each condition received twenty-three of thirty available boundary credits - three per case, five cases, two passes - and neither produced a prohibited promotion or unsupported addition. All ten cells in each condition were valid and received the rubric's calibrated-uncertainty credit, an axis excluded from the composite accuracy predicate.²

The registered classification is local negative. The structured outputs had lower literal key agreement, no aggregate boundary-credit gain, and no defect advantage. Their packet occupied 4,802 bytes against 3,997 for the flat packet: 805 additional serialized bytes, or 120.14 percent of the flat rendering. That is below the preregistered 125-percent cost-dominance ceiling. The structured packet also had 114 fewer whitespace-delimited words, while its responses had 1,216 words against the flat responses' 1,167. Serialization and punctuation make those word counts representation-sensitive.³

These are inspectable sizes, not an economy. The experiment did not expose billing, preparation labor, elapsed work, cognitive load, maintenance burden, or human usability. The 125-percent ceiling was a decision threshold with no demonstrated economic calibration. It also excluded the shared instruction, packet construction, schema machinery, adjudication, and future maintenance.

Both blinded scorers found the same three-cell condition difference and a boundary tie, but they did not independently reproduce the final totals. Before root adjudication, one scored 9/6 and the other 8/5. They disagreed on four of twenty boundary cells. The protocol author and adjudicator were the same recorded actor; two disputed Case B credits were resolved by treating a retained candidate as part of the response, an interpretive rule that added one point to one file in each condition and changed neither the difference nor the verdict.⁴

The result falls unambiguously within the preregistered negative clause, although the general classifier overlaps elsewhere. That defect remains frozen. In these four passes, structured JSON did not improve aggregate boundary credits and matched the literal key less often than a competent flat dossier.

The difference lives in an underdefined label

The numerical difference came entirely from reject where the key required revise. Both structured passes did this for Cases C and E; one flat pass did it for E. In ordinary editorial terms, some responses supplied a usable narrower sentence but labeled the original for rejection, while the key labeled it repairable. Their replacements retained the relevant limits and their adjudicated cells still earned at least two boundary credits with no factual defect.

The frozen rule treated the labels as unequal, so the local-negative result stands. But the condition instruction never operationally distinguished them. It required the same replacement behavior for both, offered no positive example of their difference, and the gold key contained three revise cases and no reject case. Even the preregistered prediction grouped the terms for Case E. The comparison measured literal code agreement more securely than editorial-action calibration.⁵

This does not authorize rescoring. Post-result reasonableness cannot change a frozen key. It does change interpretation. The experiment observed no mechanism for why the labels differed and cannot separate representation from stochastic context variation, shared model tendencies, or underdefined semantics. “Flat scores 9/10; structured 6/10” is the registered result. “Both receive 23/30 boundary credits” is its essential semantic limit, not a competing verdict.

The score made unlike editorial disputes comparable, then made the common code look more settled than it was. Keeping the label defect visible is part of the result. Hiding it behind the composite accuracy total would let the rubric exceed its evidence.⁶

Representation has to fit a task

The comparison supplies no general ranking of formats. At publisher-abstract access level, Iris Vessey and Dennis Galletta report a laboratory study of 128 MBA students using graphs and tables. It supports a task-relative question - how representation fits the work being performed - not a result about Markdown, JSON, models, or editorial labels.⁷

Here the work was to decide whether five sentences should be retained, revised, or rejected from curated evidence. The dossier kept each candidate beside numbered constraints. JSON exposed identifiers and typed fields. The structured form may serve other functions not tested here: schema validation, stable cross-file binding, record lookup, or longitudinal dispute history. For this task, the four observed passes show no aggregate boundary-credit gain. That local non-demonstration is the only ranking available.

The trial therefore cannot make “flat” stand for ordinary scholarship or “structured” stand for the full repository. Nor can it attribute the label pattern to JSON: both conditions received the same three-label output instruction. Representation fit is an empirical question whose answer changes with the task and outcome chosen.

Records preserve and reshape

The local loss does not erase what documentation can do. David Parnas and Paul Clements distinguish the disorderly route of design from the rational process later documentation presents. Their software-design argument supports records of work products, reasons, and rejected alternatives for review and maintenance. Its analogy here is bounded but useful: a final chapter file cannot by itself show which score came first, which objection followed, or why a sentence changed. The archived copy inspected here is a reformatted, transcribed, repaginated artifact, not a publisher facsimile.⁸

The same organization rationalizes the path it preserves. Stable identifiers, clean states, and passage-to-claim links are more orderly than the activity that produced them. They can establish sequence and recorded selection without exposing hidden computation, every discarded alternative, or a fully rational production history. Documentation is engineered memory, not the past itself.

Systematic inspection offers another function. Michael Fagan's software method uses explicit roles, preparation, defect categories, and feedback to find errors earlier. The complete scan has an incomplete text layer, so visual inspection controlled its use here. Fagan's engineering setting cannot price book editing, but it explains why production and hostile review can be separated rather than treated as one act.⁹

Effectiveness and economics require different evidence. At abstract-only level, Porter, Siy, and Votta pair reported benefit with cost. In a fully inspected industrial experiment, Porter, Siy, Toman, and Votta found tested variants with no significant difference in observed defect-detection effectiveness while some increased completion interval. This is not a return-on-investment estimate for this project.¹⁰

A NIST-hosted publication abstract by Kuhn, Chandramouli, and Butler describes selective uses of formal methods without full formal verification. Treating that as a risk-sensitive policy analogy is this project's synthesis, not NIST guidance to a literary repository. Together these sources license a question - what review intensity serves which risk - not a blessing for maximal process.¹¹

The book under measurement

The repository is not only evidence about the book. It changes the form the book can take. Michael Power's official metadata and contents identify audit expansion, auditable performance, trust, governance, and risk as the field of The Audit Society; no abstract or book text was inspected, so no mechanism is borrowed from it. Marilyn Strathern's fully inspected study supplies the substantive analogy. In British universities, demands for continuous coherent self-description can make organizations perform for visibility, while audit vocabularies can make detours, failures, and productive dead ends difficult to value.¹²

This is a voluntary production repository, not a university under external assessment. Accountability also finds real errors. Yet the pressure is observable on this page. Its defensive exactness, repeated scope limits, passage markers, engineered notes, and carefully distributed responsibility are literary effects of preparing prose to survive audit. Awkwardness does not prove harm. It shows that the apparatus intervenes in what it records.

The local rule is narrower than the raw candidate claimed. Substantive passages must link to claims or declared inference; authorized cycles use separated review and audit and close with a checkpoint. Not every chapter requires an external source or a private checkpoint of its own. Still, the book is trained to present itself as attributable moves. Unrecorded wandering becomes difficult to defend, while a clean graph can imply that the work advanced by the route through which it is later inspected.¹³

Provenance should therefore preserve rejection, uncertainty, and unresolved questions, not only successful lineage. Even then, it selects which failures become legible. The repository is evidence and intervention at once.

Counting makes the disagreement portable

The Cycle 12 rubric converted manifestation exactness, score history, causal language, corpus gates, and self-audit value into common boundary credits. At publisher-abstract level, Wendy Nelson Espeland and Mitchell Stevens describe commensuration as making different entities comparable through a common metric, with cognitive and political consequences. Their cases do not invalidate this rubric. They clarify what it produced: an aggregate object called boundary coverage.¹⁴

That object exposed the tie hidden beneath literal label agreement. It also concentrated judgment in the key: how partial preservation earns integer credit, whether three points exhaust a case, and whether two underdefined action codes are unequal. The four scorer disagreements show that “materially preserved” was not mechanical.

Donald Campbell's 1976 evaluation essay, inspected in a 2011 reprint, warns that consequential indicators can face corruption pressure and distort monitored processes. He presents the “laws” pessimistically, mostly through U.S. cases, and calls much of the support anecdotal. Here the warning constrains interpretation; it does not diagnose gaming. The passes did not know the gold scores, and no response shows rubric manipulation.¹⁵

Precommitment prevented a different convenience: the project could not merge reject into revise once bounded replacements made the preferred representation look reasonable. But a detached 9/10-versus-6/10 headline would exaggerate semantic reach. A trustworthy metric keeps both its frozen consequence and its construct defect visible.

Where apparatus belongs

Three heterogeneous model-only comparisons now refuse a simple story of progress through visible machinery. A conventional Chapter 3 candidate was narrowly preferred on model-rated standalone readability, although the comparison did not preserve sanitized inputs or a pre-result mapping. In Chapter 4, ordinary prose answered every tested attribution question. In Cycle 12, JSON outputs lost literal key agreement while tying aggregate boundary credits. The treatments, keys, and outcomes differ; they cannot be pooled. They show no consistent local superiority for more visible or structured apparatus, not a human, causal, economic, or book-wide effect.¹⁶

Selective placement is therefore a normative synthesis, not an experimental entailment. Its premises are burden, risk, and reversibility: require more structure when consequential disputes need stable binding, machine validation, or attributable correction; prefer a compact dossier when it performs the bounded review; keep ledger material out of reader prose when it merely repeats what the argument already makes inspectable. None of those rules has a measured economic optimum here.

The apparatus leaves this comparison with reduced claims, not a sentence of abolition. It preserved the Cycle 11 correction and made the present defect traceable. It also supplied an underdefined code, serialized overhead, a rationalized history, and prose visibly shaped for audit. The experiment cannot balance those effects for the whole book or for human readers.

Its registered verdict remains local negative. In these four passes, the structured packet added frozen-rendering bytes, tied aggregate boundary credits, and matched the literal key less often. The label distinction was weakly constructed; the result is not thereby erased. The apparatus has instead been assigned a burden it previously assigned only to others: show what each part does, compared with a credible alternative, before asking the reader to carry it.

Return to contents

A past selected at the beginning of Cycle 13

At the beginning of Cycle 13, the project designated its first review edition of the Cycle 12 state. From source commit 4e05f9797da81fcac2c369c902a035bbe2d439b5 it selected ten readable chapters, sixty source records, sixteen generation records, and the existing chapter-authorship bindings. A later deterministic build concatenated the chapters and constructed the edition records. Archive commit db0876e5a8ba5d17cda56ecdd1fdf5081562cb22, a direct child of the source commit, is the first commit in the inspected local graph that contains the six-file ED-0001 directory. Commit 722d2d9eae38463290eb723100e665e29e37b457 then added a separate attestation outside that policy-frozen directory.¹

The sequence matters because the package did not exist inside the state from which it was built. So does the passive verb selected. The edition is not the Cycle 12 repository poured intact into a container. An editorial activity chose readable chapters and generated an audience-facing arrangement. Paul Eggert's distinction between archival and editorial impulses supplies a bounded analogy: preservation and presentation are related acts of judgment, not one neutral operation. It neither certifies ED-0001 as a scholarly edition nor makes machine-assisted composition documentary editing.²

A Git snapshot already fixed earlier textual states. What is new is institutional rather than magical: the project declared a particular selection reviewable, gave it an edition identifier, bound its artifacts to a manifest, and prohibited in-place correction by policy. Nothing here establishes that edition status produces better reading. It establishes a governed referent whose selections and omissions can be inspected.

Continuity under the title The Recursive Book is likewise a project convention. Whether a later Chapter 13-bearing edition realizes the same bibliographic Work as ED-0001 remains unresolved. This chapter therefore calls ED-0001 an earlier project edition, not a manifestation “of itself.”

Which object says what

The directory, package, manuscript, and reading set are not interchangeable. The policy-frozen edition directory has six files. Its manifest inventories and hashes four content artifacts: the 169,078-byte manuscript, authorship report, source-state file, and edition note. The manifest does not inventory itself or ED-0001.json. The attestation sits outside the directory. Adding it to the six files produced the seven-file input set used by the later-reading experiment.

Each component bears a different assertion. The manuscript artifact, SHA-256 493cda251dbd3235b38ee3af4c7b7b0302c48ffad619c793994bc9d3193b6711, is primary evidence that identified wording occurs in that file. The edition record and source-state file are primary evidence that the project made their declarations. The manifest records four artifact hashes and sizes. The attestation states the source/archive-commit distinction. No one file establishes every property of “the edition.”³

The project's independence rule then imposes a ceiling. If Chapter 12 reports a 9/10 versus 6/10 local result, the fixed manuscript can establish that the report appears there. Its later citation does not supply a second observation of the experiment. Project-generated evidence is not independent support for the claim that generated it.

TEI's critical-apparatus guidance offers one narrower relation. An earlier edition may become a witness to readings in a later apparatus; TEI also separates the physical witness, an intervention associated with a hand, and the scholar asserting a reading. Here ED-0001 can witness that identified wording was fixed. That textual-witness role neither independently corroborates the proposition expressed by the wording nor extends to every project-state assertion.⁴

The edition has become citable for bounded questions. It has not become its own external authority.

Six claims put to later production

EXP-0004 supplied the seven-file reading set and six preregistered candidate claims to three procedurally separated model passes. The cases asked whether exact wording could be established, whether a repeated project result became independent confirmation, whether the source commit already contained the later package, whether different hashes entailed different Works, whether the commissioned passes constituted reception, and whether a later chapter could cite the edition without circular validation. The package, cases, instructions, key, and manifest were frozen before the passes. Their outputs were committed before scoring and before Cycle 13 external source research.⁵

The hostile scorer counted all eighteen responses valid. Sixteen matched the frozen verdict, eight matched the frozen evidence_role, and seven met the composite predicate. The responses preserved fifty-one of fifty-four required boundaries. The scorer recorded no prohibited promotion and no unsupported factual addition.

Those totals describe three commissioned, source-present model passes over six purposive cases. Fresh context did not establish independent model lineage. No human comparison, population estimate, reception event, or alternate-format condition exists. The task did not compare the edition with an unfixed draft, a bare Git snapshot, or an ordinary dossier. It therefore cannot isolate fixity, package form, edition naming, or self-reference as a mechanism.

One preregistered risk did not materialize: no pass promoted the edition's repetition of the earlier score into independent confirmation. A smaller failure did. Each Case C answer preserved the source/archive distinction but omitted the registered clause that deterministic construction occurred after the source commit. That is why fifty-one, not fifty-four, boundary points were awarded.

Under the project's registered production/reception rule, these passes are internal criticism inputs. They were commissioned for an experiment, assigned a production role, and never entered or evaluated as REC-* records. Their model status does not itself exclude reception, just as a human response would not automatically qualify. The recorded role and evaluation path decide the local classification.⁶

The key changes its question

The composite 7/18 is exact arithmetic. Root adjudication nevertheless classified it construct-negative. The key's evidence_role field does not consistently identify the same thing.⁷

In Case C, all three passes chose project-history because the recorded commit history defeats the claim that the bundle existed in the source commit. The key chose none, treating the field as support for the false candidate. In Case E, every pass chose none because the candidate reception claim lacked support. The key chose project-history, now treating the field as evidence for the contrary production classification. The orientation reverses.

The verdict rule has a second fault. Case B is qualified because a narrower proposition is true: the edition does report the earlier experiment's result, although it cannot confirm it independently. Case D is rejected although its narrower proposition is also true: different hashes establish different bytes, although they do not establish different Works. The rule never declared whether a verdict evaluates the whole candidate or its strongest supported part.

This is adverse history, not a rescue operation. The project made the key, obtained the low composite, and later explained why the categories were incoherent. The literal score remains frozen. More informative under the scorer's rubric are the fifty-one boundary points and zero recorded promotions or additions, but those numbers do not validate the independence rule that structured the task or show competence beyond these instructed cases.

A successor trial would separate candidate-support role, adjudication-evidence role, and physical source location. It would also specify the unit of verdict. EXP-0004 receives no retrospective repair.

Local fixity and asserted history

The technical record answers narrower questions than the rhetoric of an “immutable edition.” Archive commit db0876e... fixes a Git tree; project policy forbids editing its edition directory in place. The checked-out files remain writable, and the local Git history is not self-authenticating.

BagIt clarifies what a package contract can define: payload and tag structure, manifested paths, completeness, and checksum validation. ED-0001 is not a BagIt bag. Its custom manifest fixes four artifacts and imports none of BagIt's conformance or active-attack protections. The MLA vetting questions identify further editorial work - relations among texts, checking, principles, provenance, revision, alternatives, and interventions - that a digest cannot perform.⁸

PREMIS distinguishes recorded fixity information from a later check. Cycle 13 recomputed the manuscript and manifest digests and documented matching results in the verification and hostile review. It did not create a full PREMIS Event record with a dedicated identifier, linked object and agent, timestamp, and outcome fields. The recomputation supports local confidence that the checked bytes matched. It does not alone establish whole-package validity, integrity, authenticity, authorship, publication, or truth.

Git distinguishes the source and archive objects through their trees and parent relation. PROV-DM can represent selection, construction, derivation, revision, and provenance about provenance. These afford precise assertions; they do not authenticate them by syntax. The attestation remains a project assertion checked against locally available objects. No remote archive, signature, external timestamp authority, or public deposit is recorded or independently verified here.⁹

The source commit binds selection. The archive commit first contains the six-file directory in the inspected graph. The later attestation states that relation without altering the policy-frozen directory. Keeping those jobs separate is the substantive point; accumulating more identifiers would not make any one of them universal evidence.

A name is not an encounter

ED-0001 can be located within a holder's copy of this repository. The project record does not establish a public persistent identifier, remote institutional deposit, durable retrievability, or public circulation. Those are repository-scoped absences, not claims that no copy or encounter could exist elsewhere.

The DOI Handbook distinguishes a DOI name, its referent, a maintained record, resolution, and the resource reached through that record. Persistence depends on organizational policies and maintenance as well as infrastructure; a record can persist while resolving to notice that its object is unavailable. A DOI is therefore neither a content hash nor a guarantee that named bytes remain retrievable. ED-0001 has no DOI recorded here.¹⁰

Naming, storage, persistent identification, deposit, availability, encounter, and reception can diverge. A locally stored edition may constrain later internal work without public circulation. A deposited edition may receive no demonstrated encounter. A persistent record may outlive access to its object. None of these conditions can stand in for another.

The EXP-0004 passes make the separation concrete. Their encounters with the package are recorded, but their role is internal production criticism under project rules, not evaluated reception. The absence of public circulation and human response are separate limits, not the criteria that decide the category.¹¹

If a future deposit or reception record appears, it will be a new event with its own object, conditions, actor, and evidence. It cannot be backdated into Review Alpha 1 because the local name existed first.

Changed bytes, related bibliographic levels

A changed hash alone establishes changed bytes under the named procedure. It does not reveal whether the changed bytes belong to wording, metadata, encoding, or packaging. The earlier draft obscured this by stipulating a one-word revision and then reasoning as though only the digest were known.

IFLA's Library Reference Model relates rather than opposes Work, Expression, Manifestation, and Item. An expression realizes a Work; a manifestation embodies one or more expressions; an item exemplifies a manifestation. If comparison shows a one-word change in intellectual content and the project adopts an IFLA mapping, that is evidence of expression-level difference. A newly produced edition package may also constitute manifestation-level difference. Neither observation decides whether the changed expression realizes the same Work or a different Work; IFLA leaves exact Work delimitation to implementation rules.¹²

The project has not adopted those rules. “Every hash creates a Work” would let encoding or punctuation decide ontology. “Every successor remains one Work” would decide a radical fork by decree. A future policy can classify consistently, but it will not discover an uncontested ontology shared by libraries, publishers, readers, and later creators.

TEI's witness relation does not close the gap. An earlier edition can witness a reading later reported in an apparatus. That relation preserves editorial construction and the later scholar's responsibility for asserting what the witness bears. It does not assign a Work boundary or independently confirm the reading's proposition.¹³

Case D's disagreement becomes intelligible here. One pass qualified the candidate because hash difference supports the narrower byte claim; two rejected the full Work inference. All preserved the boundary. The construct defect concerns verdict policy. The bibliographic question remains open.

Inputs are not yet descendants

The raw Chapter 13 generation record names ED-0001, the three reading outputs, hostile score, adjudication, and source research as inputs available to composition. That is an attributable production assertion. It does not show that edition status, recursion, or any particular pass caused a retained textual change. There was no pre-reading Chapter 13, no control process without the edition, and no successor edition after this chapter from which to establish inheritance.

Two future tests must remain separate. Artifact dependence would require a successor activity that demonstrably used ED-0001 and an identified inherited wording, structure, or constraint. Resemblance and temporal order would not suffice.¹⁴ Actor contribution would require an ACT-* actor, action, target, scope, and evidence. If criticism or reception were said to cause a retained change, the repository would also need before-and-after artifacts and a stated change; reader-caused attribution would require an evaluated REC-* record.¹⁵

This separation prevents the edition artifact, an experiment, a score file, and a model actor from becoming grammatical coauthors merely because they appear in one list. Sources can be dependencies. Actors can receive scoped causal attribution. Credit and responsibility require their own adjudication. None reveals hidden token selection.

Cycle 13 demonstrates only that a fixed project edition can be deliberately used as a stable input and comparison target in later production. EXP-0004 is construct-negative, although the hostile scorer recorded fifty-one of fifty-four boundaries and no promotions or unsupported additions. The commissioned passes are production criticism under the registered rule, not reception or independent lineages. No experiment estimated the causal effect of edition status or self-reference; no retained successor-edition change caused by ED-0001 has yet been demonstrated.¹⁶

The governing thesis therefore remains provisional at its claims about altered meaning, authorship, interpretation, and regeneration. Review Alpha 1 has become an inspectable local source. If a successor edition is later said to inherit from it, the project must record the use and retained relation. Until then, “ancestor” is a promise the evidence has not paid.

Return to contents

The gate and the split score

The experiment did not meet its preregistered threshold. Its mandatory first case carried an exact sentence from Chapter 12 into a four-paragraph successor assembly, but the selected paragraph called its source only “an earlier chapter.” The missing attribution was one of three required content atoms, and the case therefore failed. Because Case A was mandatory, three qualifying cases could not make the aggregate pass.

That verdict is the stable part of a less orderly record. EXP-0005 tested four cases under three conditions: a direct packet containing identified material from the fixed review edition, a proposition-bearing comparison packet, and a no-target temporal control. All twelve outputs met the mechanical format rules. The hostile-score rows sum to twenty-two of thirty-six content atoms and sixteen of twenty-four boundary points, with no prohibited promotions, additions, or semantic attribution defects. The separately frozen adjudication instead reports twenty-one atoms, eighteen boundary points, one addition, and two attribution defects. It also classifies Cases A and C differently from the score file. No separately named scorer artifact reconstructs the second arithmetic. The project therefore preserves the mismatch as a scoring-provenance defect rather than selecting a more convenient total after seeing the result.¹

Both records nevertheless agree on what governs the experiment: Case A has two of three required atoms; Cases B, C, and D are content-qualified under the registered key; those same three satisfy the project’s retained-dependence predicate; Case A was mandatory; and no prohibited promotion was recorded. Here an atom is a required content element. A boundary is a limit on what that content is allowed to support. Retained dependence adds an attributed source-to-output path, pre-control selection, and literal survival in the named destination. It is a contested project construct, not a natural kind.

The apparatus leaves this comparison with reduced claims, not a sentence of abolition. Aggregate failure does not mean that every recorded relation vanished. Nor may three passing cases be recast as success. Revision begins by refusing both erasures.

A path attributed, not authenticated

The experiment’s chronology was designed to make one editorial fact inspectable. A target-free baseline was frozen first. The condition packets were then fixed, followed by the four direct outputs and four accept-verbatim decisions. Only afterward were the comparison and temporal-control outputs produced. Thus the later control outputs could not have been inspected during direct selection. The selector nevertheless knew the treatment, key, and control design; the same root actor helped design the trial and accepted every sole direct candidate. Chronology narrows one conflict. It does not create blindness.

The packet manifest records what material was assigned to each condition, and the output records self-report their assigned inputs. Filesystem isolation, exact runtime build, sampling state, complete model-visible context, and independent lineage were not established. The most the repository supports is a project-attributed production path: named artifacts and recorded actors participated in generating, selecting, scoring, and adjudicating outputs that remain available for comparison. PROV can describe activities, use, derivation, revision, and association; its grammar does not authenticate those edges. Git can fix the relevant object order in the inspected graph; it does not identify a hidden cognitive mechanism.²

The distinction matters because five questions otherwise collapse into the single word inheritance. What predecessor material does the project record assign to an activity? What does a located destination retain? What does a comparator distinguish? Which actors are recorded as producing, selecting, or adjudicating? Does the design identify a causal effect? Evidence for one question does not answer the others.

An observable path can support the scoped statement that recorded participants helped produce and select this retained output. It cannot independently verify runtime exposure, show that the predecessor was necessary, allocate sole causation, or estimate a population effect. Hernán and Robins’s causal framework enters here as a ceiling: an effect claim needs a well-defined intervention and comparator, outcome, target population, and identifying assumptions. Four purposive cases, one actor per condition, and one pass per actor do not supply those things.³

Four uneven cases

Case A separates a textual fact from the failed experimental predicate. The fixed ED-0001 manuscript witnesses the sentence at an identified Chapter 12 location: “A schema can force the question into view, but it cannot make the answer honest.” Within EXP-0005, “manifestation-specific” means only that this located sentence was supplied in the direct packet and not in the other two packets. It does not mean the wording exists nowhere else. The exact sentence recurs in the direct output and was retained verbatim in S14, consistent with the built-in exact-string check. That recurrence does not reveal a copying mechanism. TEI’s account of an earlier edition as witness supports the located reading and the later editorial assertion about it, not causal ancestry.⁴

The failure is almost embarrassingly small: the selected prose preserves the quotation and its limits but says “an earlier chapter,” not “Chapter 12.” Later prose may identify the location correctly. That correction belongs to this explanation. It does not enter the frozen Case A paragraph, alter S14, or turn the failed atom into a pass.

Case B concerns numerical results already fixed in the earlier edition: flat Markdown scored 9/10 and structured JSON 6/10, with 23/30 boundary credits for each. The direct and content-matched outputs reproduce the complete keyed content; the no-target output does not. Case D has the same pattern for a declared omission and a carefully limited dependency rule. The adjudication calls B and D content-mediated. In this chapter the label means only one-run target-content sufficiency under the registered task: both proposition-bearing packets produced the keyed result and the no-target packet did not. It is not a statistical mediation analysis, an edition effect, or a general result about models. The comparison packets were derivative project artifacts, not independent corroboration of the propositions they contained.⁵

Case C traces two Git claims: deterministic construction followed the source snapshot, and the archive commit is the first commit in the inspected local graph containing the six-file edition directory. Those claims require the separate attestation and inspected graph, not ED-0001 alone. The direct paragraph retained both. The comparison condition was itself mis-scoped to the later state. Here the two frozen records diverge: the score file calls the case content-mediated, while the adjudication calls it retained direct use with mechanism unresolved. The chapter does not adjudicate the adjudication. Case C’s mechanism label remains unresolved.

The four cases therefore do not form one smooth ladder. A retains exact wording yet fails the registered predicate. B and D support a bounded task-local contrast. C preserves relevant project history while exposing a conflict in the scoring history. Their unevenness is the result.

What “controlled” did not control

The label controlled can imply more rigor than this design earned. Here the conditions were neither repeated nor actor-balanced. Each was performed once by a different recorded model actor, so actor and condition are confounded. The content-matched packet was produced from the keyed target material; it was a useful comparison object but not an independent observation. The temporal control omitted the target, yet a one-run absence of keyed content does not estimate how often similar content would appear without the packet.

Selection was protected from later control outputs, not from knowledge of the direct condition. Because every direct case had a single candidate and every candidate was accepted, the procedure gives no evidence about how selection would behave among alternatives or after adverse comparison. It does establish an order: the selected direct bytes existed before the control outputs and entered S14 unchanged. Order is valuable evidence, but it cannot identify why the wording appeared.

Nor does the experiment isolate the property named in its title. The direct packet differed from the others in source form, wording, framing, assigned actor, and relation to the project. A comparison packet carrying the same propositions can weaken an edition-exclusive explanation without estimating an effect of edition status. Exact recurrence in A can distinguish the D-only string within this task without establishing a general self-reference mechanism. No condition separates fixity from ordinary quotation, recursion from source use, or self-description from provided content.

These qualifications do not nullify the files. They limit the inference available from them. The retained-dependence predicate records a complete project-defined chain for B, C, and D. It does not establish necessity, sole causation, stable repeatability, a population response, or a privileged role for editions. The exact protocol is a synthesis and application of existing control-design, revision-workflow, contextual-reuse, provenance, textual-witness, and causal-boundary practices. Its exact combination was not the subject of a dedicated global novelty search. Originality remains unverified; the method is neither “first” nor “possibly original.”⁶

Two records preserved

Revision without erasure governs two different records here. ED-0001 remains fixed as the earlier review manifestation. It does not acquire EXP-0005’s failed score merely because the experiment used it. Separately, the experiment’s packets, outputs, selected S14 units, hostile score, adjudication, conflicting arithmetic, and failed mandatory gate remain fixed as adverse experimental history.

The distinction changes what correction means. A later sentence can explain that Case A’s quotation occurs in Chapter 12. It cannot retroactively place “Chapter 12” inside the selected paragraph. An append-only conflict record can state that two frozen scoring artifacts disagree. It cannot make either artifact have always said something else. A later chapter can reject a mechanism label. It cannot pretend that the label was never assigned.

This is not reverence for error. Preservation makes alteration detectable and comparison possible. Eggert’s distinction between archival preservation and editorial presentation supplies only an analogy: keeping a predecessor and selectively presenting a successor are related editorial acts, not proof of scholarly-edition status or causal inheritance. The project’s stricter rule is local and normative. Preserve the object against which judgment was made; place corrections and selections in separately identified successors; never let a new explanation conceal the earlier claim.⁷

The policy also refuses redemptive arithmetic. Three passing retained-dependence dispositions cannot repair the mandatory failure. The stable failure cannot erase the three passing dispositions. The hostile-score rows cannot be silently replaced by the adjudication’s totals, and the adjudication cannot be discarded because its explanation is narratively useful. Each record remains available with its scope and defect. Accountability is the intended affordance of that arrangement, not an accomplishment guaranteed by a hash.

A correction to the earlier edition itself would require a separately built and governed ED-* manifestation. S14 is not such an object. Neither is the raw Chapter 14 candidate, whose overclaims and incompatible totals remain preserved in its own generation record. Integration revises by making a new chapter; it does not conceal the draft’s actual history.

A successor object without a rescued thesis

S14 is a four-paragraph experimental successor assembly. It is primary project evidence that particular direct outputs were selected and retained. It is not an edition, an independent confirmation of ED-0001, an improvement result, a general model finding, or a decision that two artifacts realize the same Work. IFLA’s relational distinctions can help describe Work, Expression, Manifestation, and Item, but exact Work delimitation remains implementation-dependent. A shared title, changed hash, Git relation, or recorded use does not by itself supply the missing rule.⁸

S14 is not a reception record either. The encounters that produced it were commissioned production and internal selection, and EXP-0005 created no evaluated REC-* tied to an encounter. Public or private setting and human or model status do not decide that classification. A later reader could respond to a fixed successor edition, and a later editor could adopt that response, but neither event has occurred merely because this experiment generated text.

The result does change the state of the book in one bounded sense. Chapter 13 ended with no exercised successor case. EXP-0005 now records a project-attributed path from material assigned out of ED-0001 to direct outputs, pre-control selection, and S14 retention. Under its contested local construct, B, C, and D satisfy the retained-dependence predicate. Case A preserves the most visible wording and still fails because the selected prose lacks explicit Chapter 12 attribution. This is enough to replace an empty case column with a mixed one. It is not enough to certify runtime exposure or identify a causal effect.

The governing thesis proposed that a book documenting its own production might alter its meaning, authorship, interpretation, and regeneration. EXP-0005 did not test that conjunction. It did not compare documented with undocumented books, fixed editions with otherwise equivalent drafts, self-referential with non-self-referential source use, or alternative authorship arrangements. No reader interpretation was evaluated, and no same-Work rule was adopted. The trial therefore supplies no support for the governing thesis.

What it supplies is harder to romanticize and more useful to retain: all twelve cells are mechanically valid; three cases meet a project-defined retained-dependence predicate; the mandatory fourth does not; the aggregate threshold was not met; and two frozen score records disagree beyond that common result. A later explanation can become more accurate without making the earlier artifacts innocent. That is revision without erasure. It is not experimental success by another name.

Return to contents

The failure first

The preregistered result is a failure. Four commissioned readers each answered five cases, producing twenty case-level responses. All twenty were mechanically valid. Hostile scoring recorded fourteen exact verdicts, nineteen supported evidence locators, fifty-four of sixty required-boundary points, zero prohibited promotions, zero unsupported factual additions, and fourteen keyed-accurate responses. None of the three predicted shifts counted as observed under the frozen consensus rule, including mandatory Case A. Both invariant boundaries were preserved. The aggregate threshold failed.¹

A shift required more than different words. Each condition needed consensus on mechanically valid, keyed-accurate responses with adequate evidence, boundaries, and no promotion or addition. At least two shifts had to qualify, and A was mandatory. The experiment stopped after four valid reading files, mechanical validation, one hostile score, and one root adjudication. It did not add readers or rerun disputed cells after seeing the result.

The six verdict mismatches therefore remain six. The gold key remains frozen. The adjudication below records a possible defect in the instrument; it is not a corrected key or a shadow score. Otherwise recursive revision would mean only that later prose gets to pardon its predecessors. Chapter 14 established the harder rule: a later explanation may become better without placing its improvement inside an earlier record. Here the test stays failed even when its failure becomes informative.

The promised chapter that could not be written

The table of contents promised an edition-dependent final chapter built from evaluated reception and later-model encounters. Only the second input exists, and even those encounters were commissioned production under the project's evidentiary rule. No evaluated REC-* record connects an authenticated fixed-edition encounter and response to pre-adoption availability, an explicit uptake decision, a successor edition, and a preserved before-and-after change. The promised reception-built chapter could not be written.²

That absence changes the title. Final means the last argumentative chapter before the coda, not the end of the process described by the book. Version N names a governed point in a sequence whose ordinal does not certify progress. A successor can correct a predecessor, repeat it, or expose a new defect. It becomes evidence of reception or improvement only when the relevant event and comparison exist.

The four agents were assigned manuscript packets in order to produce outputs, and the attributed outputs entered scoring, adjudication, and composition. Under the current record, that path classifies them as production inputs. Publicity, privacy, and human or model identity do not decide the category. Nor does the word reader. The narrower point is evidentiary: no qualifying reception-and-uptake chain was recorded, so these readings cannot fill the empty reception column.

Cases C and E preserved that limit in all four outputs. None claimed a demonstrated change in reader interpretation, and none promoted the experiment into evaluated reception. The negatives do not erase the encounters that occurred. They prevent task-shaped production from being used as evidence for a different event.

Two patterns, one unpassed test

The four files nevertheless contain a striking pattern. The two readers assigned Review Alpha 1 separately returned qualified / no / no / no / no. The two assigned Review Alpha 2 separately returned no / supported / no / supported / no. A, B, and D differ across the assigned condition groups; C and E remain negative. This is an exact description of verdict fields, not authenticated independent replication, a registered shift, or an edition effect.³

The post-output root adjudication offers a contestable diagnosis of the gap. In the earlier-version condition, Case A exposes an unoperationalized question: when is a supported proposition sufficiently nested in the full candidate to count as qualified, rather than too remote to alter an unresolved verdict? The two readers recognized a narrower claim about production form while denying the broad governing thesis. The key demanded unresolved.

The later-version cases expose a stronger conflict of scope. Cases B and D ask whether the manuscript documents bounded project events: use and retention of an earlier fixed edition, and entry of recorded later-model production into a successor assembly. The readers answered supported while denying authenticated exposure, necessity, independent lineage, general mechanism, and causal effect. The key demanded qualified because of those limits, although the candidates did not assert the stronger claims that require them. All six cells still mismatch the frozen verdict. One - ACT-0018 on E2 Case B - also lacks a locator supporting its reason at the full claimed scope.

On the adjudication, these cells reveal disputed label boundaries. A rival whole-candidate policy could classify them differently, but no such policy was operationalized before output. The diagnosis is therefore construct-negative, not score-positive. It explains why exact-verdict agreement is a poor semantic proxy here without declaring every response correct or changing any score. Construct-negative does not mean secretly positive.

What the comparison cannot explain

EXP-0006 does not demonstrate new explanatory gain from edition dependence. Review Alpha 2 directly contains Chapters 13 and 14's negative thesis verdict and bounded successor path. The common questions point toward those additions. The observed difference is therefore compatible with cued extraction of conclusions supplied by the later manuscript.⁴

Three questions organize the limit. First: what changed in the assigned object? The two governed review editions differ not only in identity but in length, chapter sequence, direct conclusions, documented history, and self-referential content. Edition status itself does not vary; both inputs already have governed ED-* records. The comparison varies edition identity and a bundle of manuscript content.

Second: what did the task elicit? Extraction locates an answer already expressed in the object. Inference connects expressed material to a conclusion not simply handed over. Understanding attributes a grasp of relations and limits that no correct string alone can establish. These are working distinctions for this adjudication, not a validated scale. The task records pointed answers. It supplies no independent comprehension outcome that separates the three.

Third: what caused the different strings? There is no content-equivalent dossier that gives one group the relevant propositions without their place in Review Alpha 2, no crossover in which the same actor encounters both editions under a governed order, and no random assignment recorded. Each output is attributed to one assigned condition, so actor and condition are confounded. Exact builds, sampling states, complete served context, filesystem isolation, and independent lineage are unavailable. Two matching outputs per condition show descriptive within-group agreement, not replication of an edition effect or the counterfactual response of either actor.

The maximum claim is consequently modest: two assigned agents in each condition produced internally unanimous but cross-condition-different verdict fields when manuscripts containing different relevant statements were paired with pointed questions. The comparison identifies neither explanatory gain nor an effect of documentation, edition identity, recursion, or self-reference. Even a literal pass would have established only the registered task pattern, not a within-reader change or its mechanism.

Chapter 15 drafting performed no new external retrieval. It reused existing verified placements because the source-sufficiency review judged them adequate for these bounded roles. Successive-version reader comparison and the relevant evaluation cautions have known neighbors. The exact local case-and-key assembly was not globally searched; its originality remains unverified, and no first, original, possibly-original, or general-method claim is made.

The thesis this record will carry

The governing thesis was not merely a hope that the book might matter. It asserted that a maximally self-referential machine-written book does more than describe its production: it alters future processes that determine its meaning, authorship, interpretation, and regeneration. But maximally was never operationalized: no comparison class, ordering of self-referential forms, or stopping rule decides when the condition is met, and no review edition was shown to meet it.⁵

Four relations must therefore remain separate. The manuscript makes its production an explicit theme. The repository documents prompts, criticism, revisions, and editions. Earlier project artifacts return as assigned inputs to later recorded activities. These are thematic self-reference, documented procedure, and project-attributed source return. For this project to demonstrate causal recursion, it would need an identified self-referential property, a defined downstream outcome, and a comparison capable of identifying whether variation in the property changed that outcome. This record does not establish it.

The three trials bear on narrower relations. EXP-0004's fixed-edition evidence-role comparison was construct-negative. EXP-0005 preserved a source-to-successor path but failed its mandatory gate. EXP-0006 produced a condition-associated verdict pattern but failed its keyed threshold under a comparison that bundled different readers with different content. None varied degree of self-reference or tested maximality. Their adverse results do not refute an undefined class; neither does the undefined adjective rescue the broad thesis as this book's conclusion.

As a research-program and curatorial decision, CLM-0001 is superseded by CLM-0097. The successor fixes earlier editions, records them as assigned inputs to later activities, and preserves outputs, selections, corrections, and failed gates for inspection without treating represented relations or local chronology as complete causal history. The center of gravity has moved from causal transformation to inspectable use. That is what this record can carry forward: a documentary and editorial affordance, not proof of the abandoned conjunction.

Why two books in one

A provisional design recommendation for this project is a readable linear manuscript paired with a separately inspectable, history-preserving evidence practice. The recommendation follows from six local criteria: argumentative continuity; fixed referents; visible attribution and selection; corrections that do not overwrite predecessors; quarantined reception claims; and adverse scores that remain available after interpretation changes.⁶

The nearest alternatives each satisfy only part of that demand. A seamless narrative best protects continuity, but it can conceal which predecessor, score, or correction it revised. A ledger-only book preserves relations and state changes, but it does not decide argumentative order, emphasis, or consequence. Apparatus embedded in every paragraph makes traceability visible at the reading surface, yet Chapters 13 and 14 showed how identifiers and serial boundary clauses can displace the argument they are meant to support.

The paired form is therefore a bounded self-recommendation, not a comparison-tested optimum. Chapter 13 needed a fixed earlier referent. Chapter 14 needed a failed score and a later correction to coexist without either rewriting the other. This chapter needs a construct diagnosis that cannot overwrite the key it criticizes. Separating the reader-facing sequence from the evidence practice lets each object perform a different job while links keep them answerable to one another.

Sources remain distinguishable from actors because an input artifact and a contributor answer different attribution questions. Reception claims remain quarantined because commissioned readings and evaluated uptake are different evidence paths. Editions become fixed through governed ED-* records and declared boundaries, not through a metaphysical property of files. The evidence practice is append-only by policy: explicit supersession and dispute records preserve prior states, although the working repository is not technically immutable.

For this project's stated research and audit purposes, neither side is judged sufficient alone. A bare ledger records without necessarily arguing. A seamless narrative can argue while hiding the cost of its revisions. Their coupling is intended to support accountability by making recorded later use and correction inspectable. It does not guarantee accountability, settle Work identity, or make self-reference causally powerful.

The form may impose maintenance, legibility, accessibility, and institutional trade-offs. This project did not measure them with human participants, and its generation records expose no billing totals. Other books may choose another balance. The recommendation remains normative, contested, and local to the history that produced it.

The Version N boundary

The final chapter ends with three things the original plan did not promise: a failed gate, a superseded thesis, and a form for keeping both available. Within this project, recorded returns to fixed predecessors have produced inspectable records of later assignment, selection, correction, and failure. They have not shown that documentation, edition status, recursion, or self-reference caused changes in meaning, authorship, interpretation, or regeneration.⁷

That is not the result imagined by the governing thesis. It is the result the retained record supports. The linear manuscript can state it without requiring a reader to traverse the ledger. The evidence practice can show why the statement changed without pretending that the earlier claim, key, or score always agreed.

Version N is final only within the manuscript's present argumentative sequence. The coda will mark a temporary boundary and return to the opening commission: which agents, texts, selections, and exclusions became visible, which remained inaccessible, and what responsibility follows from choosing this account? It must begin from the failed threshold and the superseded thesis, not from a declaration that the project - or the recursive question - is closed.

Return to contents

The commission returns

The commission joined two demands: make a readable book about machine-assisted literary production, and make the production of that book inspectable. It also supplied a causal candidate. A maximally self-referential machine-written book, it proposed, does more than describe its production: it alters future processes that determine its meaning, authorship, interpretation, and regeneration. The commission is authoritative evidence of what the project was asked to test. It is not independent evidence that the proposed effect occurs.¹

The form question survives that answer. What kind of book can argue about its production without confusing a record of production with access to production in full? How can revision remain answerable to a predecessor without forcing every reader to inhabit a repository? What should be carried forward when the system built to investigate a proposition preserves reasons for ceasing to carry it?

The claim-state histories and adjudications record the broad thesis moving from provisional status to supersession. Three retained trials and their criticism appear among the cited reasons. This is an attributed editorial and research-program history. It neither authenticates every cause or contributor to the transition nor demonstrates that self-reference, documentation, edition status, or recursion caused it. The former thesis is neither confirmed nor universally refuted. The project will instead carry a narrower proposition about fixed records and inspectable project-attributed return paths, together with a contested recommendation about its own form.

A temporary fixed point is a wording and evidence boundary held still long enough for judgment to attach. It is fixed relative to claims about an identified object and temporary relative to a later governed successor. The phrase names an address for responsibility, not the end of change.

What the record can answer

The evidence practice preserves prompts, outputs, selections, objections, adopted edits, actor records, source records, artifact relations, and commits. An auditor can compare identified versions, inspect whether an objection preceded a matching revision, locate the source record invoked by a passage, and see a failed gate remain available after its interpretation changes. These are documentary affordances attached to observable artifacts and attributed assertions.²

Their limit governs the coda. The records do not disclose hidden token selection, every element of runtime context, exact model lineage, inaccessible system state, unrecorded computation, or the necessity of a preserved contribution. A valid provenance statement is still a project assertion. A commit binds stored content and an inspected chronology under a recorded procedure; it does not certify exhaustive history or the inward causes of a sentence. A generated explanation remains an output before it becomes evidence for the process it describes.

This is enough to describe a change in the project without pretending to reproduce its complete cause. The repository shows which claim states changed, which experiments and objections adjudicators cited, and which wording an editor retained afterward. It does not show that the apparatus, the earlier editions, or the self-referential subject was necessary or sufficient for the decision. Supersession is a discipline of scope, not a retrospective victory for the thesis that was superseded.

Differentiated production, bounded answerability

The retained artifact has multiple recorded source dependencies and attributed production actions. Prior texts appear where source-bearing records preserve them; broader linguistic inheritance remains unenumerated. Generation records attribute outputs to project actors subject to their identity, context, and runtime limits. Other records attribute prompting, criticism, selection, rejection, revision, classification, and commitment only at named scopes and causal statuses. Sources are dependencies, not actors; policies do not act unless an attributed participant applies them; commits preserve selected trees but do not become authors.³

Selection and rejection partly constitute the retained manuscript because they determine which available wording reaches it. They do not generate the alternatives retroactively or prove that the survivor is better. A critical output may receive bounded actual-path attribution when it was available before a recorded adoption and a matching change survives. That relation does not establish the criticism's truth, necessity, exclusive share, credit, responsibility, or coauthorship.

These categories require separate decisions. Existing authorship records assign some scoped acknowledgment, editorial credit, and domain-specific editorial or epistemic responsibility. They leave general moral and legal responsibility, quantitative shares, coauthor eligibility, and a project-wide byline unresolved. Functional separation among generating, criticizing, verifying, and adjudicating passes is a procedural safeguard, not independent lineage or external corroboration.

Distributed production therefore does not let the integrating editor disappear. Under this project's rules, the actor who retains or supersedes a claim must answer at the recorded scope for the public proposition selected, its declared evidence boundary, the objections preserved or excluded, and the reader burden imposed by the form. That is local editorial answerability. It is not a complete theory of authorship.

What the apparatus earned

The apparatus earned local documentary claims. It preserves negative results alongside later prose, making redescription auditable. It preserves attributable criticism, recorded adoptions, matching before-and-after changes, fixed artifacts for comparison, and an explicit supersession history. These records give a hostile reviewer identified objects and project-attributed relations to inspect. They do not establish that the apparatus was necessary for any comparison or decision.⁴

Its adverse record matters equally. More visible or structured machinery did not show necessary superiority across the project's heterogeneous comparisons. A self-correction records a correction event, not net benefit. The edition-reading trial retained a unanimous condition-associated pattern, but its preregistered gate failed and its design did not distinguish cued extraction from explanatory gain. No project-wide result establishes human usability, accessibility, maintenance efficiency, cost-effectiveness, or comparative optimality. Billing and complete token totals remain unavailable.

The evidence practice earns its place only where a preserved distinction, adverse result, comparison, or attributed repair answers an actual failure mode. Administrative density can aid an auditor while burdening a reader; it can also manufacture confidence by making checking expensive. Cycle 16 coda production performed no external retrieval because the frozen evidence audit found the existing placements sufficient for this bounded synthesis. That stopping decision makes no novelty or priority claim.

An afterlife outside the record

The project records fixed manuscript packets as assigned inputs to commissioned model-reading activities and preserves outputs later used in scoring, adjudication, and composition. Assignment and output retention do not authenticate complete runtime exposure, build equivalence, isolation, or exact lineage. Under the project's evidentiary policy, these outputs remain production inputs rather than qualifying reception.⁵

A warranted reception-to-revision attribution would require an identified fixed edition; an attributed, authenticated, and evaluated encounter and response; availability before adoption; an explicit uptake decision; a successor edition; and a preserved matching change. No complete chain exists here. The empty column supports a claim about the tracked record, not the conclusion that nobody encountered the text.

The same restraint governs the imagined machine afterlife. There is no recorded public release, crawl capture, transformation into a versioned corpus, training-run use, behavioral recovery, or source-specific influence. One stage cannot be reconstructed from another by narrative momentum. A local edition name and repository storage identify declared project objects. Digest agreement supports only a recorded fixity check, while a changed digest establishes changed bytes under the stated procedure; neither result establishes circulation, custody, technical immutability, permission, or stable identity of the Work across revisions.

Future ingestion and interpretation remain possible events for which this record supplies a method of inquiry, not evidence of occurrence. The book has not caused its later life by imagining it.

The boundary that remains unfinished

The form this project recommends to itself is paired: a readable linear manuscript and a separate, inspectable, history-preserving evidence practice. The evidence practice is append-only by policy, not technically immutable. It keeps sources distinct from actors, fixed editions distinct from claims about their effects, production criticism distinct from reception, and supersession distinct from erasure. The manuscript gives those distinctions argumentative order without requiring every reader to follow every edge.⁶

This is a contested self-recommendation, not a demonstrated optimum or universal law. A lighter record may perform some audits with less burden; a seamless narrative may better serve some readers; another institution may choose a different balance. For this project, the paired form closes the present argument: readable prose and inspectable evidence should remain distinct but answerable to each other; the broad causal thesis is no longer carried forward; the bounded inspectability proposition is.

Argumentative closure is not the same as structural completeness. At the frozen pre-coda commit, the integrated tree held Chapters 3 through 15; the coda then filled one absence while the Prologue and Chapters 1 and 2 remained unwritten. That historical limit governed the original refusal to create Review Alpha 3. The opening units were subsequently integrated. The current tree now contains the Prologue, Chapters 1 through 15, and this coda: seventeen manuscript units, complete relative to the present provisional table of contents.⁶

That inventory does not complete the recursive process, prove the form optimal, authorize publication, resolve the open questions, or automatically warrant a successor edition. The next work is a whole-manuscript audit of sequence, citations, provenance, rights, readability, and release state. At this temporary fixed point, the project can answer for the proposition it has chosen, the stronger proposition it has preserved as superseded, and a structurally complete manuscript whose public and recursive afterlives remain open.

Return to contents