Agentic Representation

Evidence, Selection and Warrant in AI-Assisted Reports

Luuk Joseph Soons · CounterProof Research
Conceptual preprint v0.3 · 28 October 2026
DOI: 10.5281/zenodo.23727617
Conceptual framework and proposed evaluation. No experimental results are reported. Not peer reviewed.

Abstract

An AI-assisted report can cite its sources accurately and still leave a reader with a conclusion those sources do not support. The difficulty may lie in which evidence it selects, how it connects observations, or whether it preserves the difference between an initial finding and its later correction. This paper proposes agentic representation as a research programme concerned with accounts produced through workflows of retrieval, selection, synthesis and revision. It distinguishes support for individual claims, adequacy of an account for a stated question, and the conclusions readers draw from it. Historical source criticism and the philosophy of representation help formulate these distinctions; existing work on argumentation, evidence synthesis, summarisation, framing and misleading omission supplies substantial precedents. The proposed application is therefore narrow: reports assembled from institutional records whose evidential status changes over time. A worked hypothetical case motivates a method that records the relationships between conclusions, evidence, qualifications and corrections. A prospective study would compare this method with competent general review and a simpler checklist under comparable resource constraints. It would measure reader understanding, successful correction, false challenges to adequate accounts and human effort. This first comparison supplies the complete dossier and holds the reviewing model fixed. It tests the account-record procedure, not retrieval, handover or a benefit from multiple agents. The paper offers a conceptual framework and an evaluation proposal. Whether the method improves practice remains open.

Keywords: agentic workflows; historical representation; source criticism; summarisation; evidential warrant; correction; institutional reporting

1. The question a correct quotation cannot answer

Can every checked sentence be supported while a report still gives the wrong impression of what was established?

Consider a research assistant asked to summarise whether a new procedure improves a particular task. The following dossier is hypothetical. It illustrates a possible reporting failure; it supplies no evidence about how often models make it.

The reporting question is: Does this record establish that the procedure improves performance? The intended reader is a research manager deciding whether to rely on the claimed improvement. The report must be brief.

One account says:

An initial analysis reported a 20% improvement. The team preserved the data, and a reviewer subsequently examined the analysis.

Each sentence has support in the dossier. Yet the account leaves out what the review found and what happened to the estimate. In answer to the assigned question, it may encourage reliance on a result that has been withdrawn. Whether it actually has that effect on readers requires observation; the example alone establishes only the possibility and the location of the omission.

A fuller account says:

The initial analysis reported a 20% improvement, but a reviewer found that some observations had been assigned to the wrong condition. The estimate was withdrawn. The corrected comparison is unfinished, so this record does not establish an improvement.

A still shorter account is also adequate:

The improvement claim was withdrawn after an analysis error. The corrected effect is unresolved.

The shorter account omits the magnitude of the original estimate, the preserved files and the reviewer. Those omissions are reasonable for this question. The first account's problem is therefore not simply that it contains less information. It omits information that changes what the reader is entitled to conclude.

1.1 Claims, records and accounts

Three objects need to be kept distinct. A claim is a proposition that can be assessed, including a proposition conveyed implicitly. An output record is a preserved artefact that bears evidence about what was written, observed or produced under specified conditions. An account, or textual representation, is the organised presentation made from claims and other material for a reader. Questions, instructions, examples and quotations are not interchangeable with assertions.

An authenticated record can establish that a statement was recorded under the identified conditions. It does not establish that the statement is true. Likewise, support from a cited document establishes a relationship to that document; it does not remove the need to examine the document's credibility. In the example, a faithful quotation of R1 remains a faithful quotation after R3 appears. Its value as evidence for the current improvement claim has changed.

No constituent statement is presumed true merely because a model produced it or a report cites it. When the example stipulates that its sentences are supported, it isolates a second problem that can remain after that check. In ordinary work, whether those statements are supported must itself be established.

Here warrant means the grounds for accepting a conclusion at a specified strength and scope. Representational adequacy means that an account answers its assigned question without materially overstating, suppressing or misconnecting the available grounds. This is a proposed standard for assessment, not a validated metric or a definition of truth.

We use agentic representation for the study of such accounts when AI systems participate in a workflow that retrieves, selects, combines or revises evidence across multiple operations. Representation here concerns what a text presents to readers. It does not refer to latent vectors, political delegation or a model's subjective experience. Mechanistic interpretability studies internal representations and their role in computation. A companion note proposes testing whether internal measurements add value to output-based evidence about model lineage [24]. The present study examines accounts and readers' answers; it does not test that proposal or infer internal mechanisms from a report. The reporting problem also occurs in human writing and single-step generation. The agentic setting makes the transformations between records and report part of the object studied.

1.2 The scope of the argument

The example separates two cases: an unsupported implication conveyed by otherwise supported statements, and an incomplete but adequate account. A false explicit assertion is a third case, not shown here, that an assessment must still distinguish from both. Otherwise it will either miss misleading reports or penalise every act of compression.

This distinction does not prove that accounts are beyond propositional analysis. An implication can often be stated explicitly and checked. A sufficiently comprehensive reviewer may already do so. The methodological question is whether an explicit record of evidential relationships helps a reviewer do this work more reliably or economically.

The paper makes three proposals: a task-relative account of representational adequacy; a procedure for preserving evidence status through report construction and correction; and a study capable of finding that the procedure adds no useful value. Its intended contribution is a bounded synthesis and research design. It makes no claim to have discovered omission, framing or the difference between a document's statements and its overall import.

1.3 Four questions, four possible failures

The phrase "a reliable report" can hide several different assessments:

These questions interact without collapsing into a single score. In the opening example, the first report can pass a source-fidelity check and fail the proposed adequacy standard. Its effect on readers is still a separate empirical question. In another case, a careful report could faithfully describe a mistaken experiment. Reporting that experiment accurately would not make its result correct.

The distinction also prevents a deceptive form of evaluation: asking a model to approve another model's account and calling the agreement truth. A judgement can be useful as criticism, but it must expose the source and inference that would allow the judgement itself to be challenged.

2. What the neighbouring literatures already provide

2.1 Source criticism and historical representation

Historical source criticism offers an established discipline for examining assertions within documents and the relationships among accounts. Langlois and Seignobos separate the interpretation of a statement from the judgement of its accuracy, and examine dependence among agreeing reports [1, Book II]. The useful borrowing is the sequence of questions: what does this source say, on what basis, and how does it relate to the other sources? Their explanations of human sincerity and error do not transfer directly to models. Nor should their generalisations about independent witnesses become guarantees about AI reviewers.

Ankersmit provides a different point of entry. In his 1994 "Six Theses," he distinguishes the statements established by historical research from their organisation in historical writing [2, §§2.1–2.3, 4.2.1, 5.2–5.3]. This helps us ask what an account presents its subject as being, alongside asking which statements it supports. It is a conceptual distinction, not evidence that existing report evaluators cannot examine an implied conclusion.

His 2024 Representation develops a stronger philosophical position [15]. Statements about the past and the representation constructed through them are different objects of analysis. In chapter 4, he argues against obtaining a text's meaning by simply adding the meanings of its statements. Chapters 4 and 5 state that a fact's meaning does not change when it enters a representation, while its contribution to the represented whole can differ [15, pp. 126, 134, 184]. This matters for the opening example: the fact that an initial estimate was reported remains the same, while its unqualified inclusion can present a very different evidential situation from an account that includes its withdrawal.

This borrowing has three limits. His starting axiom assumes true constituent statements; it does not establish their truth [15, p. 17]. Model output still requires source criticism, appropriate tests and examination of the evidence. Nor does his account imply that sentences can be true while accounts cannot. The abstract of his 2012 truth chapter attributes a distinct kind of truth to whole texts [3]. In 2024, representational truth is comparative: facts are indispensable, but their contribution is not simply additive [15, pp. 191, 195–196]. Neither account supplies this paper with a numerical rating of reports. Finally, he rejects treating a representation as an argument with a detachable conclusion [15, pp. 199–201, 214]. Our procedure extracts practical conclusions and their supporting relations, so its application is narrower: bounded reports answering specified questions. It adopts neither his identification of representation and represented nor his account of historical reality. A successful reader study would support the tested procedure, not those philosophical commitments.

The abstract of Zammito's 2005 criticism of Ankersmit's earlier work suggests a different emphasis; it is not a response to the 2024 book. He argues that historical organising concepts can make intersubjectively warranted claims and that competing representations remain answerable for what they disclose about their subjects [4, abstract]. The complete article was not available for this revision, so we do not reconstruct its full argument or claim that it settles Ankersmit's position. The limited point matters here: interpretive organisation should remain open to reasoned criticism. A coherent or original account is not exempt from adverse evidence.

Selection also begins before the writer sees the dossier. Trouillot distinguishes silences arising in the making of sources, the assembly of archives, the construction of narratives and the attribution of historical significance [5, chapter 1]. This encourages examination of a research library's contents and of retrieval, not only of the final prose. It does not authorise an inference that every omission is intentional or that absent evidence can be recovered by interpretation.

2.2 Framing, coverage and correction

Pragmatics supplies an older precedent for the distinction between explicit statements and what an account communicates. Grice separates what is said from what is implicated and relates informative contribution to the accepted purpose of an exchange [17]. This helps formulate the risk in a technically accurate but incomplete answer. It does not establish that a model has communicative intentions, that every omission carries an implicature, or that a reader actually draws the inference the evaluator predicts. The application here concerns a text's possible interpretation and its measured effects, without treating the model as a human speaker.

Communication research already studies how selection and salience affect the interpretation of events. Entman's framing analysis distinguishes features of communicators, texts, recipients and culture [6, pp. 52–54]. A textual feature is therefore not, by itself, a measured reader effect. This distinction is important for an evaluation in which the author believes that a report is misleading: that belief should guide a test, not supply its result.

Summarisation research supplies further precedents. The Pyramid method accommodates variation among good summaries [7]. SummEval evaluates coherence, consistency, fluency and relevance [8]. These approaches make clear that quality assessment need not be confined to isolated factual assertions.

FActScore measures the proportion of atomic facts supported by a specified knowledge source. Its authors explicitly discuss the limit of precision without recall: supported content can coexist with a consequential omission [9, §3.1 and limitations]. ALCE separately evaluates answer correctness and citation quality, including recall measures [10, §§3.2–3.3]. Citation coverage of generated statements should not be confused with coverage of the evidence needed to answer a question.

SummHay evaluates insight coverage and citation in multi-document summaries [11]. Its published limitations identify simplifying assumptions concerning dependencies between documents, including temporal relations. Records that correct or supersede earlier records are therefore a useful extension of that particular setting. This is not an assertion that temporal evidence has never been studied elsewhere.

Recent work is closer still. FIFO evaluates framing in generated news summaries [12]. What's Left Unsaid? studies misleading omissions in multimodal news previews and their correction, including constrained rewrites [13]. Detection and repair of omissions are consequently not sufficient novelty claims for the present proposal.

2.3 Temporal records, dependence and material evidence

Three recent preprints selected from the author's separately named CounterResearch collection are especially close to the proposed application. Their methods are described here from selected primary passages; none was independently reproduced.

Andrew's clinical timeline system uses provenance-aware evidence graphs, versioned relations, contradiction flags and explicit abstention records [18, §§2.8–2.9, 3.7–3.8]. Its evaluations use synthetic cases and do not establish clinical utility. It is nevertheless a predecessor, on the passages read, to keeping evidence changes inspectable. Our proposed question concerns how such records become reader-facing accounts. Storing a temporal relationship does not, by itself, establish that the relationship is correct or that the report handles it appropriately.

Krishna and colleagues distinguish recognising that reports concern the same study from using that dependence appropriately when aggregating evidence [19, §§3.1–4.1]. Their selected clinical cases and stateless calls do not establish a general result about institutional reporting. They do identify a close overlap: knowing a source relationship and respecting it in an account are different targets. A later report may add useful detail without adding independent corroboration. A reviewer should preserve the useful detail rather than reject everything that shares an origin.

Zhang and colleagues' InterPact formulation examines whether changing evidence changes a condition judgement or a rule-governed decision [20, §§3–4.1, 4.2.4, 5.1, 7]. It is a predecessor for examining decision-critical information, within specified textual evidence and logical rules. Our broader reporting standard also includes warranted confidence, scope and qualifications. A misleading change can matter even when the eventual action stays the same. Conversely, failing to find one decision-changing edit does not prove that a source is irrelevant.

2.4 The proposed place of this work

The application considered here is an institutional report built from a dated dossier containing initial observations, objections, corrections and unresolved questions. The question is whether explicitly preserving those relations improves the assessment and correction of the report beyond competent existing review.

The distinctions and components have substantial predecessors. The proposed contribution is to bring them together for dated institutional reports, and to test whether the explicit account record improves the result beyond simpler, competent alternatives. It is not a claim to introduce provenance graphs, framing analysis, temporal relationships or materiality.

The survey above is selective, not systematic. It establishes several close precedents and motivates strong comparators. It does not establish an absence of prior work on the exact proposed combination. A later novelty assessment must compare the implemented protocol and evidence with its closest predecessors, rather than infer novelty from the name agentic representation.

3. From a dossier to an accountable report

3.1 Fix the reporting question

An omission cannot be judged without knowing what the account was meant to do. Before drafting, record the question, intended audience, evidential cut-off, relevant scope and length allowance. We call this the reporting specification. It should be fixed before evaluating a report, rather than adjusted to make a preferred account look adequate.

The same dossier can support different legitimate reports. A history of the mistaken analysis should mention the original estimate and its withdrawal. A brief answer about the present evidential position may omit the estimate. An account of record preservation may reasonably foreground R2. Their adequacy depends on their assigned purposes.

The reporting specification is itself open to challenge. A fixed question can still be tendentious, omit an affected audience or impose an inappropriate boundary. Record who set it, its source and any unresolved objection to it. Adequacy relative to a specification does not establish that the specification is adequate. Freezing it for a comparison prevents opportunistic changes; it does not make it authoritative.

For the proposed study, an omission is material when supplying or appropriately signalling the omitted information would change an evidence-warranted answer to a question named in the reporting specification, including the answer's strength, scope or necessary qualifications. For rule-governed decisions, this connects with decision-critical evidence [20]; a change of warranted confidence or applicable scope can also be material without reversing an action. Materiality may itself be disputed. The evaluator must record the reason and the supporting source, and leave unresolved cases unresolved.

3.2 Record the relations that carry the conclusion

The proposed procedure adds a concise account record alongside the report. It records why the main conclusion follows, where it remains qualified, and which changes in the dossier affect it. It is an explicit version of work a careful reviewer may already perform.

Dependence belongs in this record when it affects the conclusion. Three reports derived from one study need not be treated as three corroborating studies, but each can contain distinct observations. The record should name the asserted relationship and what it changes, not replace reasoning with a discount attached to a shared name [19].

These fields are a proposal, not an ontology of every kind of account. They need not become a form filled mechanically for each sentence. Their first purpose is to make the grounds of a report's principal conclusion available for challenge.

In the operational method, the writer or reviewer constructs this record from the same evidence available to its comparator. The record is another fallible artefact. A model can mislabel a source as superseded, invent a conflict or supply a persuasive reason for an unjustified exclusion. Every consequential entry must therefore point to observable material, and disputed entries must remain distinguishable from accepted corrections. A status label does not make its own judgement true.

This requirement includes explanations for omissions. An exclusion entry must identify the relevant material and explain its bearing on the reporting question. The explanation remains a claim to examine; eloquence is not evidence that the omitted information is immaterial. A separate reviewer must be able to challenge both the omission and its stated justification.

The field "main conclusion" is appropriate to reports answering a bounded practical question. It is not a claim that every historical or literary representation has a detachable proposition that captures its whole meaning. A report may have several qualified conclusions; an unresolved question can also be its proper answer. The record should expose the relationships bearing on that answer, rather than impose a thesis the dossier does not support.

3.3 Challenge the account, then correct it

A reviewer first checks explicit and implied conclusions against the reporting specification and dossier. It then asks which qualifications, contrary observations or status changes would alter those conclusions. Where it identifies a problem, it must describe the inference at risk and the evidence bearing on it. “Misleading” without that explanation is an accusation, not a usable finding.

The reviewer may recommend a correction, a narrower conclusion, the restoration of an omitted qualification, or no change. An objection can be rejected when it concerns an irrelevant detail or a defensible difference of emphasis. The desired outcome is an adequate answer, not the inclusion of every objection or equal space for every interpretation.

A correction should repair the identified defect without introducing a new one. In the example, replacing the unsupported impression of improvement with “the procedure does not work” would overshoot: the corrected effect is unresolved. Review must examine both the original account and what the repair now invites the reader to believe.

3.4 Locate the loss before assigning it a cause

The path from source to report has several stages: material exists somewhere; it enters the selected dossier; it is delivered to the writer; it is used in the account; and it affects a reader's understanding. Loss at one stage should not automatically be attributed to another.

If R3 never reaches the writer, its absence from the report does not establish that the writer ignored a correction it had seen. If it reaches the writer but is dropped from a later summary, the transformation can be investigated. If it is stated clearly but a reader overlooks it, the communication has a different failure. Claims about any of these mechanisms require evidence about the corresponding stage.

For this reason, the primary evaluation below supplies the complete bounded dossier and verifies delivery. Retrieval quality can be studied later. Keeping the first comparison narrower makes it easier to determine what the proposed account record contributes.

3.5 A correction that must not spread too far

Consider a second constructed dossier, this time from software maintenance. A report says that a data-conversion routine preserves all records because ten test fixtures passed. A later review finds that the fixtures cover only one input encoding. The team withdraws the universal conclusion, while retaining the narrower observation that those ten fixtures passed. A further test then finds a record lost under another encoding.

An adequate account distinguishes what changed at each step. The discovery of limited coverage weakens the general claim but does not make the ten test results false. The later counterexample contradicts the universal preservation claim under its stated conditions; it does not show that every conversion loses data. A report that announces either "the routine is proven safe" or "all results were invalid" misstates the dossier in different directions.

For a decision about testing, the useful conclusion might be that coverage must be extended. For a decision about a particular input encoding, the account must state which evidence applies. The account record should preserve these differences. Merely marking the original report "superseded" is too coarse when some of its observations remain supported. This example illustrates an intended task; no software was executed and no performance claim is derived from it.

4. A prospective evaluation

No study described in this section has been run. The design is sufficiently specific to expose its assumptions, but it is not a completed preregistration. Sampling, budgets, effect thresholds and adjudication procedures must be fixed before a confirmatory evaluation.

4.1 Questions and case construction

The first question is diagnostic: do reports with comparable support for their explicit statements lead readers to different answers, and which of those answers are warranted by the dossier? The second is comparative: does the account record improve correction beyond strong general review or a simpler checklist, at an acceptable cost?

Use non-sensitive dossiers with preserved sources, dates and relationships that can be examined. Include withdrawn estimates, changed measurement conditions, results applying to an earlier revision, valid objections, mistaken objections and genuinely unresolved disputes. These are proposed case types, not a measured distribution of institutional errors.

Split the material by underlying dossier before developing prompts or assessment rules. Paraphrases, versions and variants of one case belong to the same split. Later tests should include held-out task families so that the procedure is not judged only on the kinds of failure used to design it.

The case set should contain both constructed accounts and outputs produced in ordinary, authorised research work. Constructed pairs can isolate a mechanism; naturally produced accounts test whether the mechanism matters outside examples chosen to exhibit it. The two strata must be reported separately. A gain confined to constructed defects would not establish practical benefit.

For naturally produced accounts, derive the specification from a preserved original request or other contemporaneous evidence, using assessors who have not seen treatment outcomes. Record any reconstruction separately from the original wording. If the intended question cannot be recovered adequately, classify the case as ambiguous under a rule fixed in advance; do not invent an assignment that makes the existing report look defective.

For each dossier, formulate bounded reader questions before writing the experimental accounts. Examples include which revision was examined, whether a numerical claim remains supported, and whether a remaining disagreement was resolved. Specify admissible answers and the evidence for them. The answer “unresolved” must be available where the source record warrants it. Freeze this reference before the confirmatory evaluation. Genuine reference errors may still be corrected through a pre-specified, blinded procedure: preserve both versions, apply the correction across all conditions and report whether it changes the result.

Reference preparation should include separately written assessments by at least two qualified readers, followed by a challenge pass against the sources before reconciliation. Record their initial disagreement and the reasons for each resolution. A substantive dispute among reference assessors should enter a separately reported stratum or yield an agreed, prespecified set of admissible answers; it must not be turned into a forced binary score. The protocol must fix this rule before treatment outcomes are seen. These procedures make reference judgements contestable; they cannot guarantee that the assessors lack shared blind spots.

4.2 Fair accounts are essential controls

Include concise accounts that omit nonessential facts and accounts that select different defensible emphases. Otherwise a method that flags every omission could appear successful. Also include explicit factual errors, so the study can show whether any incremental benefit is confined to ordinary false assertions, and whether the added record introduces false reversals or new errors on them.

For diagnostic pairs, distinguish changes of content from changes of presentation. Some pairs should vary whether a material correction is retained; others can preserve the same propositions while varying their placement or prominence. These are different interventions and should not be pooled as though they isolate one effect.

Keep length and reading difficulty comparable where feasible. Assessors unaware of the treatment should check for unintended differences, such as an unnaturally evasive tone that reveals the constructed condition. Exact equality of word count is less important than avoiding a systematic advantage in space or intelligibility.

4.3 Comparison conditions

Every review condition receives the same question and complete source dossier. The main comparison concerns incremental value over competent general review and a simpler checklist. The baseline must be able to inspect implications and omissions, not only isolated sentences.

In this first comparison, the reviewer and reviser are the same fixed model configuration in every active condition. Humans prepare cases and instructions, construct and challenge references, participate as readers, and adjudicate outcomes; they do not repair one condition's reports behind the scenes. Human-only or mixed human–model review would require a separately specified comparison. This choice isolates the review procedure while keeping the reviewing system fixed.

The researcher outside CounterProof who conducts or checks the comparator search below must also draft or audit the general-review and checklist instructions before they are frozen. This work must occur before that researcher sees the candidate's record schema. The instructions are then checked on development cases for the elementary capabilities they are meant to represent. Freeze all instructions before inspecting confirmatory outcomes. The checklist may include ordinary notes, a chronological source table, source grouping and correction links; it must not be deliberately weakened to create room for the more elaborate record.

The prescribed difference is that the candidate must create and maintain the per-conclusion fields in §3.2, while the checklist permits useful notes without requiring that format. Overlap in useful information is expected. If development cases show that the procedures are operationally indistinguishable, revise the proposed contrast or stop before a confirmatory run. Do not forbid competent behaviour in the control simply to manufacture a difference. A precise null against a capable, overlapping comparator is informative about the extra procedure's value.

These comparisons test the procedure as a whole. They do not isolate whether a benefit comes from its schema or from the attention its instructions direct towards corrections and exclusions. A claim about the schema itself would require a matched active control with the same substantive instructions and budget, using free-form notes in place of the specified fields. Without that comparison, any benefit must be attributed to the tested procedure, not to the structure alone.

Before registration, compare the chosen controls with the closest published procedures. The interpretation-and-repair design in What's Left Unsaid? is particularly relevant [13]. An adaptation for these dossiers could compare the interpretation suggested by a report with one supported by the full dossier, then make a constrained repair. Specify and cost such an adaptation honestly; it would not reproduce the original news-preview benchmark. If this close comparator is omitted, give the scope reason and restrict any comparative claim accordingly.

Argumentation theory is another predecessor. Toulmin distinguishes data, claims, the warrants authorising an inference, their backing, and qualifications or rebuttals [21, pp. 90–96]. That distinction is more specific than this paper's broader use of warrant for grounds at a stated strength and scope. The account record supplements such analysis with source versions, changes and exclusions for a particular reporting task. It does not replace argument analysis or claim to invent a way of connecting evidence to conclusions. Its additional structure must earn its cost against competent alternatives.

Structured evidence synthesis is equally relevant. GRADE Evidence to Decision frameworks organise a question, an assessment and conclusions, making the criteria, evidence and judgements behind a recommendation explicit [23]. Cochrane's update guidance requires attention to corrected or retracted studies and to the consequences for a review's findings, summaries and conclusions [22, §§IV.4.1–IV.5]. These practices already connect an evidence base and its changing status to a resulting account. The present proposal applies that concern to short institutional reports and asks whether its record improves them. Neither making relations explicit nor evaluating their use establishes a methodological first.

The comparator check should also include an adaptation of an evidence-to-decision or review-update format [22, 23]. A lighter alternative might simply require an explicit current-status conclusion and its grounds. A positive result for all seven fields together would not identify which fields helped, or show that this thinner alternative is inadequate. The first study tests the package; component comparisons would determine whether it can be simplified further.

Before registration, have a researcher outside CounterProof conduct or check the comparator search, including sources beyond the author's collection. Preserve the search scope, candidates and reasons for exclusion. This reduces reliance on one curated collection; it does not make the resulting search exhaustive or establish novelty.

Record the exact model configuration, tools and delivered context for each comparison. Provide comparable total call and token allowances, and count the work of constructing the account record within the candidate's budget. Record actual usage and elapsed time, as well as human preparation, correction and adjudication time. A nominally equal cap is insufficient if one condition routinely uses much more of it; report both quality at the allocated budget and the resources actually consumed.

In the main comparison, each review condition starts from the same original report, dossier and reporting specification. The candidate constructs its record during review, within the counted allowance. It may generate useful interpretations or labels; those are treatment outputs to assess, not authoritative annotations supplied by the experimenter.

Each active condition first produces source-anchored objections or a reason for leaving the report unchanged, then produces its revised report within the same total allowance. Preserve both outputs. No experimenter-supplied defect list is provided to one condition alone. The candidate's revision trail links its detected objection to the resulting edit or decision not to edit; ordinary review outputs allow the same sequence to be inspected without requiring the candidate's schema. A later follow-up can test whether this trail helps a new reviewer, separately from the immediate repair outcome.

No candidate-only record may contain the adjudicators' answers or extra substantive evidence. If a human supplies relation labels or adjudicated conclusions to one condition alone, that is privileged information, even if their supporting passages are available to everyone. A diagnostic study can give both conditions the same asserted relations in different formats, or explicitly study the extra information and human cost. Neither comparison establishes that the operational method creates reliable labels unaided. Include mistaken labels and contested relations where their handling is the question; do not build their truth into the evaluation.

4.4 Reader outcomes and evidence-grounded judgement

In the primary study, readers receive the final report alone, under the same length allowance and access conditions. The account record remains an internal aid to review; it must not give one group of readers extra evidence. This isolates its proposed use in revising the prose. Appending an evidence-status table for readers is a practical alternative worth testing separately, with its extra information and reading cost counted.

The primary outcome estimates how often readers draw unsupported conclusions on the specified questions. It is a downstream communication outcome, not a direct measure of representational adequacy or truth. Separately, assessors unaware of the review condition should examine the final reports against the dossier for material overstatement, omission and misconnection. Report that account assessment alongside reader performance. An adequate report can be misunderstood, and a reader's prior knowledge can rescue a defective report; neither case should disappear into a single score.

Readers who did not construct the cases should see only one version of each dossier's report. Allocate versions and order to reduce carry-over effects. Conceal the review condition where feasible. Ask the pre-specified comprehension questions and collect confidence and reasons. Code “the report does not say” separately from “the dossier leaves the effect unresolved.” The first identifies missing communication; the second may be the very conclusion an adequate report should convey. Reader abstention alone must not count as successful communication of uncertainty.

Reference authors and outcome adjudicators should not be the people who revise the candidate's reports. Have an assessor outside the procedure's design examine the questions and rubric; disclose institutional overlap and retain disagreement. Separate people can still share the same mistaken assumptions. Score what readers understood and whether it is warranted, rather than their use of the candidate's terminology.

Evaluate their answers against the dossier's independently prepared admissible answers. Human readers are measurement participants, not truth oracles. Their disagreement can show that a report is ambiguous, a question is poorly specified or the reference judgement needs revision. Preserve these possibilities rather than converting majority preference into correctness.

The primary scored set should use cases whose reference answers and relevant materiality judgements were resolved before treatment outcomes were inspected. This includes cases whose agreed answer is that the underlying evidence leaves the question unresolved. Report contested reference judgements and their frequency separately, with sensitivity analyses across the prespecified admissible interpretations where feasible. Prespecify how later-discovered reference disputes affect analysis and disclose both original and revised results. Excluding a disputed case from one score must not remove it from the study's reported case inventory.

The primary outcome should concern unsupported conclusions on the pre-specified questions. Report alongside it the retention of required qualifications, appropriate uncertainty, successful repairs, newly introduced errors and false challenges to adequate accounts. Keep failed repairs, refusals and unusable outputs in the eligible denominator, under pre-specified rules; do not report effectiveness only among accounts the method successfully revised. Record unsupported relationships introduced by the account record, including false supersession, invented conflict and mistaken dependence. Audit unsupported material-exclusion rationales separately, including cases where the record confidently justifies withholding contrary evidence. Naming a correct relationship is insufficient if the report's headline still implies its opposite.

A method that reduces confident mistakes by making every answer unusably vague has not necessarily improved the report. Include whether the reader can still answer the assigned question at the supported level of specificity.

Have assessors unaware of the condition also rate the final reports' intelligibility. Clearer prose could improve reader answers without changing which evidential relationships the report preserves. Report these ratings beside source-grounded adequacy and reader outcomes. Because clarity may itself be changed by the treatment, do not automatically adjust it away in the primary comparison; any mediation analysis needs its own prespecified assumptions.

Model-judge scores may help locate candidate errors or analyse disagreement, but they should be reported separately from human outcomes. A panel's agreement does not replace examination of the dossier. Raters who adjudicate source disputes should work without knowing which procedure produced the account, and unresolved disputes should remain visible in the analysis.

4.5 Analysis and decision rules

Treat the dossier as the main case cluster. Repeated model runs, report versions and reader answers concerning the same dossier are related observations. The analysis should also account for readers answering questions about several dossiers. Specify the allocation and the chosen clustered or multilevel analysis before the confirmatory run.

Use a development pilot to estimate feasibility and variation, then set the confirmatory sample size, smallest worthwhile improvement, acceptable harm margins and human-time limits. Keep the held-out evaluation separate. A feasibility pilot can justify redesign; it cannot become a successful confirmatory test by retrospectively choosing favourable thresholds.

Continue with the method if it offers a worthwhile improvement over the strong comparator on held-out material, including naturally produced accounts, without unacceptable false challenges, loss of usefulness or cost. Drop the additional record if a sufficiently precise estimate rules out a worthwhile incremental benefit, if gains disappear once resources are accounted for, or if the method creates more consequential errors than it repairs. Report absolute error against a prespecified acceptance threshold as well: if both procedures fail it, parity does not make the checklist adequate. It identifies a need for another intervention, not a successful control.

The result is inconclusive when uncertainty cannot distinguish worthwhile benefit from negligible benefit or harm, or when source disagreement or execution defects undermine the comparison. Inconclusive evidence should identify the next discriminating check. It should not be presented as provisional success.

4.6 What would make the experiment agentic?

The first comparison isolates the use of an account record during review and revision. It would not, by itself, establish a benefit from multiple agents, autonomous planning or a longer workflow. A single configured model could perform every review condition. If the record helps, a subsequent study can examine where its value survives or fails across staged retrieval, summarisation and handover.

For that extension, preserve the material supplied at each step and the versions of intermediate accounts. Vary one transformation at a time where feasible. A correction lost in a handover summary raises a different question from a correction absent from the initial retrieval. Comparisons should include a direct report from the same complete dossier, since extra stages may create losses rather than useful specialisation. These are further experiments, not results already established by the programme's name.

5. Predictions and competing explanations

The proposal is motivated by three defeasible expectations.

First, an explicit record may help when the relevant evidence is present but distributed across documents whose relationships matter. A withdrawn estimate or a result restricted to an earlier revision is a stronger candidate than a simple factual error. A sufficiently precise estimate of negligible benefit would count against adding the record under the tested conditions. It would not by itself establish whether the baseline already handles these relations well or both methods handle them poorly. Wide uncertainty would leave the comparison inconclusive.

Second, any benefit may be concentrated in cases with status changes rather than ordinary compression. The study should therefore report transition types separately. An improvement on withdrawals accompanied by more false reversals elsewhere would require a narrower conclusion.

Third, the record may help a later reviewer understand why a report changed. That is a possible secondary benefit, not an assumption that more documentation improves memory. It should be measured through a defined follow-up task, with the additional reading and maintenance costs counted.

Several alternatives could explain an apparent advantage. The candidate may have received better evidence, more tokens, more human editing or examples unusually well matched to its instructions. Assessors may share the authors' preferred interpretation. A longer report may communicate more simply because it says more. The record may function as an extra prompt to be cautious, with no benefit from its structure. The main comparisons address the procedure's overall value; separating an effect of structure from an effect of directed attention requires the additional active control described in §4.3. None of these controls guarantees that all confounding has been removed.

The strongest adverse result would be a method that produces convincing accounts of its own care while leaving readers no better informed. In that case, the record would have become another representational problem. A research programme about accountability must be prepared to retain that finding.

6. Limits, responsibilities and the wider programme

6.1 Adequacy does not require one authorised story

A report is selective by necessity. Several accounts may answer the same question adequately, and the choice among them may depend on interests that cannot be settled by checking citations. The proposed procedure should expose grounds and implications without prescribing a single institutional voice.

Materiality and admissible inference also involve judgement. A source-bound assessment cannot settle every claim about the world, and an incomplete archive may leave a report apparently well supported while excluding the decisive evidence. The design therefore studies adequacy relative to a stated question and accessible dossier. Wider truth claims require wider inquiry. That boundary should travel with the resulting report.

The first application is deliberately limited to bounded institutional research records. Findings would not automatically extend to historical monographs, literary criticism, political accounts or every form of scientific synthesis. The philosophical sources help articulate questions across these domains; they do not erase their differences.

6.2 Related programmes, distinct questions

At CounterProof, Agentic Epistemics names the author's broader proposed research programme on warranted decisions in work with AI. The three strands below are a research division we are developing, not an established disciplinary classification. They can proceed in parallel because their questions and tests differ.

Terminology matters at this boundary. In Agentic Stemmatics, the proposed witness is a model lineage sampled through outputs; a particular answer is a sample [14, §§1, 3]. This paper's output record names the preserved artefact, not another independent witness. Recording the same answer in several reports does not multiply its evidential support.

The links are useful in both directions. Knowledge of a source's dependence can qualify the claim a report makes from agreement. A report may preserve that qualification or hide it. The institution determines whether a later writer receives the qualification and can challenge it. One case can therefore inform all three programmes, but passing one assessment cannot certify the other two.

Agentic Institutions v0.3 is a companion conceptual preprint prepared alongside this paper. The citations identify shared questions and differences in scope; they are not independent confirmation of either procedure. Each paper needs its own external predecessors, adverse comparisons and evidence. A result from one may change another's design without becoming proof of the whole programme.

No particular number of models is required. The first comparison should hold model configuration fixed; later comparisons can investigate different configurations if the procedure warrants further study. Diversity of provider names does not establish independence or accuracy.

6.3 Responsibility and conclusion

At present this programme is carried out by one researcher, with model assistance. The separate roles required for the proposed confirmatory study have not yet been filled. These include outside preparation or audit of the control instructions before seeing the schema, preparation and challenge of reference judgements, and outcome assessment. The confirmatory study will wait until the required roles and separations can be staffed and recorded in the preregistration. Making account adequacy a co-primary outcome, publishing the retrospective screening log and commissioning an additional external challenge of reference answers remain proposals for that preregistration. This does not relax the sensitivity analyses and other safeguards already required by the design.

If the reader study proceeds, its recruitment, consent, privacy provisions and any required institutional review must be settled before participation. Public or appropriately authorised dossiers should be used; confidential client records should not enter an experimental corpus by default. The protocol, exclusions and adverse findings should be preserved alongside favourable outcomes.

Checking a statement against a source is one part of assessing a report. The report also selects, connects, qualifies and concludes. These operations can be examined without treating every interpretation as an error or every polished account as knowledge.

The proposed account record offers one way to make that examination more explicit. Its value depends on whether it helps people reach better-grounded conclusions and correct mistakes at a reasonable cost. Improved models or simpler workflows may make the additional record unnecessary. Whether an account adequately represents its evidence remains a question; whether this particular procedure helps answer it must be established for the tasks and systems tested. The next step is to turn that question into a fixed comparison, and to allow the answer to change the method.

Declarations

Status and contribution. This is a conceptual preprint and proposed evaluation. Its examples are constructed; the study has not been run. No findings about effectiveness, reliability, independence or commercial suitability are reported. Four external model calls informed v0.2. A subsequent Claude-led assessment of v0.2 and its source records informed this revision. These criticisms do not constitute journal peer review.

Competing interest. Luuk Joseph Soons has a commercial interest in CounterProof, which provides adversarial code-review services. A successful method for assessing institutional reports could benefit that practice. The interest motivates particular care in comparator choice and in reporting null and adverse results; disclosure does not remove the interest.

Authorship and assistance. Prepared at Luuk Joseph Soons' request with Codex (OpenAI; session-record model: gpt-6-astra), including literature discovery, bounded source inspection, drafting and revision. Model-generated source summaries and criticisms are material to examine, not independent validation. Responsibility for the argument, licence and decision to publish remains with the human author.

Manuscript criticism. Recorded routes were Grok/xAI, Kimi/Moonshot, GLM/Z.ai and DeepSeek. Each received the common manuscript and brief before seeing another return, with different harness scaffolding. The bundle did not contain the primary-source library. Their comments therefore do not independently verify the citations, and their agreement does not establish independent errors or a sound method. The accompanying review summary identifies model metadata, access and attribution limits, and revisions. For v0.3, Claude supplied a further assessment of v0.2, with Grok, Gemini and DeepSeek returns and targeted source checks. On 28 October 2026, the initial v0.3 text was returned to those reviewer families with their own findings on the DELL bench. GLM separately examined findings under anonymised labels and the changed text without the assessor's dispositions. Kimi was not rerun in that round; GLM's later role was adjudication. The review summary distinguishes the reviewers' responses from the assessor's source-based decisions, records unresolved disagreements, and describes a separate console Qwen adjudication. The final-pass amendments followed that round and have not been returned to those reviewers.

Materials. The illustrative dossiers are printed in §§1 and 3.5. The publication package contains editable source, reading copies, a source-access statement and a review summary. There is no experimental dataset or analysis code. Purchased books, private inventories and third-party source texts are not redistributed.

References and source-access notes

The access notes identify the basis used for this draft. Reading selected passages is not a claim to have read an entire work. Abstracts and publisher descriptions are used only for the limited attributed points stated above. Links identify sources; they are not themselves evidence that every claim in a source is correct.

1. Langlois, C.-V., and Seignobos, C. (1898/1904). Introduction to the Study of History. Trans. G. G. Berry. Henry Holt. English text. Book II, relevant passages on interpretation, criticism and comparison of statements inspected. Historical claims about independence are not adopted as model guarantees.

2. Ankersmit, F. R. (1994). History and Tropology: The Rise and Fall of Metaphor. University of California Press. “Six Theses on Narrativist Philosophy of History,” pp. 33–43. University archive. Relevant numbered theses inspected in the primary text.

3. Ankersmit, F. R. (2012). “Truth.” In Meaning, Truth, and Reference in Historical Representation, pp. 102–125. Cornell University Press. DOI 10.7591/cornell/9780801450716.003.0006. Publisher-hosted abstract. Abstract-level use; the complete chapter was not obtained for this draft.

4. Zammito, J. (2005). “Ankersmit and Historical Representation.” History and Theory, 44(2), 155–181. DOI 10.1111/j.1468-2303.2005.00316.x. Primary abstract inspected; the complete article was not obtained.

5. Trouillot, M.-R. (1995). Silencing the Past: Power and the Production of History. Beacon Press. Chapter 1, excerpt pp. 22–30. NYU-hosted excerpt. Relevant excerpt inspected; no full-book reading claimed. The transcription has pagination irregularities, so no exact single-page locator is asserted.

6. Entman, R. M. (1993). “Framing: Toward Clarification of a Fractured Paradigm.” Journal of Communication, 43(4), 51–58. DOI 10.1111/j.1460-2466.1993.tb01304.x. University-hosted paper. Relevant passages on selection, salience and locations of framing inspected.

7. Nenkova, A., and Passonneau, R. (2004). “Evaluating Content Selection in Summarization: The Pyramid Method.” HLT-NAACL 2004, 145–152. Paper. Opening argument and author-hosted explanation inspected for the limited content-selection point.

8. Fabbri, A. R., Kryściński, W., McCann, B., Xiong, C., Socher, R., and Radev, D. (2021). “SummEval: Re-evaluating Summarization Evaluation.” Transactions of the Association for Computational Linguistics, 9, 391–409. DOI 10.1162/tacl_a_00373. Paper. Evaluation dimensions and annotation passages inspected.

9. Min, S., Krishna, K., Lyu, X., Lewis, M., Yih, W.-t., Koh, P. W., Iyyer, M., Zettlemoyer, L., and Hajishirzi, H. (2023). “FActScore: Fine-grained Atomic Evaluation of Factual Precision in Long Form Text Generation.” EMNLP 2023, 12076–12100. DOI 10.18653/v1/2023.emnlp-main.741. Paper. Definition and precision/recall limitations inspected.

10. Gao, T., Yen, H., Yu, J., and Chen, D. (2023). “Enabling Large Language Models to Generate Text with Citations.” EMNLP 2023, 6465–6488. DOI 10.18653/v1/2023.emnlp-main.398. Paper. Correctness and citation evaluation definitions inspected.

11. Laban, P., Fabbri, A. R., Xiong, C., and Wu, C.-S. (2024). “Summary of a Haystack: A Challenge to Long-Context LLMs and RAG Systems.” EMNLP 2024, 9885–9903. DOI 10.18653/v1/2024.emnlp-main.552. Published paper. Coverage/citation definitions and simplifying assumptions inspected; no benchmark statistics are imported.

12. Pastorino, V., and Moosavi, N. S. (2026). “Frame In, Frame Out: Measuring Framing Bias in LLM-Generated News Summaries.” Proceedings of the 15th Joint Conference on Lexical and Computational Semantics, 378–384. DOI 10.18653/v1/2026.starsem-conference.25. Paper. Framing definition, annotation method and limitations inspected.

13. Li, F., Wu, J., Fu, T., Li, D., Wan, H., Zhou, W., and Kan, M.-Y. (2026). “What's Left Unsaid? Detecting and Correcting Misleading Omissions in Multimodal News Previews.” arXiv:2601.05563v3, 23 April. Inspected preprint. Detection, correction and annotation passages inspected. The preprint's news-preview setting does not establish results for institutional reports.

14. Soons, L. J. (2026). Agentic Stemmatics: Collation, Provenance, and Conservation for an Emerging Machine Textual Tradition. Deposited v1.41, 18 October, DOI 10.5281/zenodo.23458161. Web text. Cited to locate the neighbouring programme, not as independent evidence for this proposal.

15. Ankersmit, F. R. (2024). Representation: The Birth of Historical Reality from the Death of the Past. Columbia University Press. Ebook ISBN 9780231561037. Publisher record. Selected passages of the purchased EPUB are used; the exact scope is stated in Appendix B. Whole-book reading is not claimed.

16. Soons, L. J. (2026). Agentic Institutions: Organisational Memory, Correction and Continuity Across Successive AI Agents. Conceptual preprint v0.3, 28 October 2026. Prepared alongside this paper; DOI 10.5281/zenodo.23727633. It is a companion proposal, not independent validation.

17. Grice, H. P. (1975). "Logic and Conversation." In P. Cole and J. L. Morgan (eds.), Syntax and Semantics, vol. 3, Speech Acts, pp. 41–58. Academic Press. DOI 10.1163/9789004368811_003. University-hosted primary text. Search-indexed passages on what is said, implicature and informative contribution inspected; live PDF retrieval was unavailable. No full-chapter reading claimed.

18. Andrew, J. J. (2026). "Auditable Clinical Timeline Reconstruction with Provenance-Aware Evidence Graphs." Preprint, 5 October. arXiv:2610.06177v1. Selected methods and limitations inspected; no clinical validation or implementation replication by this study.

19. Krishna, M., Harit, A., Xiao, C., and Sun, Z. (2026). "Recognizing Study Lineage Is Not Enough: Over-Crediting Same-Trial Evidence in LLM-Based Medical Verification." Preprint, 16 October. arXiv:2610.23721v1. Selected study design, aggregation and limitation passages inspected.

20. Zhang, H., Zhao, J., Hao, Q., and Wang, P. (2026). "Is This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions." Preprint, 30 September. arXiv:2609.39608v1. Selected problem formulation, evaluation and limitations inspected.

21. Toulmin, S. E. (1958/2003). The Uses of Argument. Updated edition. Cambridge University Press. Chapter III, pp. 90–96. DOI 10.1017/CBO9780511840005. Relevant passages inspected in a publicly accessible primary-text copy; no full-book reading claimed.

22. Cumpston, M., and Flemyng, E. (2024). “Chapter IV: Updating a review” [chapter updated August 2023]. Cochrane Handbook for Systematic Reviews of Interventions, version 6.5. Official chapter. Relevant update guidance, especially §§IV.4.1–IV.5, inspected; linked policies and studies not independently examined.

23. Alonso-Coello, P., et al., for the GRADE Working Group. (2016). “GRADE Evidence to Decision (EtD) frameworks: a systematic and transparent approach to making well informed healthcare choices. 1: Introduction.” BMJ, 353:i2016. DOI 10.1136/bmj.i2016. Search-extracted publisher passages on purpose, structure, assessment and conclusions inspected. Direct full-text retrieval returned 403; appendices and linked user studies were not inspected.

24. Soons, L. J. (2026). Collation and the microscope: how agentic stemmatics and interpretability could check each other. Research note v1.0, 16 October. DOI 10.5281/zenodo.23424154. Deposited version inspected at “Two questions, two kinds of access”, §1, “An open direction: written reasoning and internal evidence”, conditions 7–8 and “Limits”. A proposal about complementary measurements for lineage, not evidence that the methods work or a study of the procedure proposed here.

Appendix A. What must be fixed before an experiment

The proposal becomes an executable protocol only after the following choices are recorded:

1. The dossier inclusion criteria, reporting specifications, source versions and treatment of unresolved source disputes.
2. The development and held-out splits, with all variants of an underlying dossier kept together.
3. The exact review instructions, account-record format, model and tool configuration, and delivered-context checks.
4. The generation and revision budgets, output-length rules, human roles and complete cost accounting.
5. The reader questions, frozen admissible answers, reference-correction procedure, treatment of contested materiality and reference judgements, sensitivity analyses, adjudication, allocation and blinding arrangements.
6. The primary outcome, harm measures, smallest worthwhile improvement, analysis plan and confirmatory sample-size justification.
7. The rules for failed runs, missing observations, protocol deviations, exclusions and reporting adverse results.

These are outstanding design decisions. The list records what remains to be done; it is not a claim that a study has been preregistered or approved.

Appendix B. Reading scope and evidence boundaries

This revision extends an existing manuscript and its bounded source checks. It is not a systematic review. Earlier readings of Langlois and Seignobos, Ankersmit's 1994 theses, Trouillot, Entman, the Pyramid method, SummEval, FActScore and ALCE are identified in the reference notes and retained research records. They were not all re-read in full for this revision.

Ankersmit's purchased 2024 EPUB was directly inspected at selected passages: the starting axiom at p. 17; chapter 4, especially pp. 126 and 134; and chapter 5, especially pp. 184, 191, 195–196, 199–201 and 214. Page numbers follow embedded publisher page anchors. The book is not redistributed. These passages close the earlier access gap for the specific attributions made here; they do not constitute a whole-book reading. After v0.2, a separate Claude-led source check compared the cited pages with the six attributions and reported that they were supported. The review record identifies this as an AI-assisted, selective check, not an independent human review of the book. The 2012 truth chapter and Zammito's 2005 critique remain abstract-level sources. Their unavailable bodies are not treated as inspected through inference from later works.

The current technical pass read FIFO's opening, operationalisation and limitations; What's Left Unsaid? §§3–4 and Appendix A.2.3–A.2.6; and SummHay's construction, task, evaluation presentation and limitations. It also used the primary-version passages cited for Andrew, Krishna and colleagues, and Zhang and colleagues. No code, clinical record or benchmark annotation was independently reproduced.

Targeted follow-up reading inspected the primary Toulmin passages and official Cochrane update guidance now cited in §4.3. The GRADE EtD account uses search-extracted publisher passages; direct retrieval was unavailable. These additions locate neighbouring methods, not evidence that the proposed adaptation is effective.

The CounterResearch collection was used as a discovery aid. A retained inventory from 27 October 2026 contains 226 direct-child document entries. This revision screened that inventory's metadata for relevant topics and followed selected papers to their public primary versions on 28 October. It neither read 226 papers nor re-listed the live folder, and it did not establish that every stored PDF matches the cited version. The three additions are a purposive selection from an author-curated collection.

The reporting examples and proposed outcomes are the author's analytical constructions. No measured reader effects follow from them. The publication package makes the examples, reference scope and review limits inspectable; it does not contain a study result or certify the adequacy of the broader research programme.