Agentic Representation: An Introduction
How a report can quote accurately and still give the wrong impression
About ten minutes to read.
1. The missing sentence
A team tests a new procedure. Its first analysis reports a 20% improvement. The team saves the data and analysis files. A reviewer then discovers that some observations were assigned to the wrong condition. The improvement claim is withdrawn. The corrected comparison is not yet finished.
This is a hypothetical case. It illustrates a possible problem; it tells us nothing about how often AI systems produce it.
Now imagine a manager asking an assistant: “Does our research establish that the procedure works better?” The assistant replies:
An initial analysis reported a 20% improvement. The team preserved the data, and a reviewer subsequently examined the analysis.
Every sentence has support in the records. Yet the answer omits what the reviewer found. It may leave the manager relying on an estimate that the team has withdrawn. The example identifies an omission; how actual readers respond would need to be studied.
A shorter answer can do better:
The improvement claim was withdrawn after an analysis error. The corrected effect is unresolved.
That answer does not say the procedure is useless. It says the available record does not establish an improvement. A later corrected analysis might find a benefit, harm or no meaningful difference.
Agentic Representation begins with this distinction. Checking that sentences match their sources is necessary work. We also need to examine what the report presents those sources as establishing.
2. What a report asks us to believe
An account is a presentation assembled for a reader: a research summary, an incident report, a review of a proposed change. Its author selects material, connects observations and decides what needs emphasis. These choices help the reader understand the subject. They can also obscure what is known.
Four questions therefore deserve separate attention:
- Does the report convey its sources accurately?
- Does its conclusion follow from the available evidence, with the necessary qualifications?
- Is the underlying claim actually true?
- What does the reader understand the report to mean?
A quotation can be accurate even when its source is wrong. An account can fairly describe a mistaken experiment. A well-supported report can still be misunderstood. Success at one of these tasks does not settle the others.
The paper uses warrant for the grounds supporting a conclusion at a particular strength and scope. In ordinary language, this means asking what we have reason to believe, how strongly, and about which conditions. A report should preserve those boundaries.
There are older traditions to draw on. Historical source criticism asks what a document says, how credible it is and how it relates to other documents. Frank Ankersmit's work draws attention to the difference between individual statements and the account formed by organising them. The paper borrows this distinction without adopting his whole philosophy or assuming that AI-generated statements are true. Its procedure concerns bounded reports answering specified questions; it does not claim to capture everything a historical work means.
Humans face the same problem. “Agentic” identifies the setting studied here: AI participates in several operations, such as retrieving, selecting, summarising and revising evidence. A qualification can disappear in one summary and remain absent from later reports built on it. Following these changes matters; machine writing does not acquire a different standard of truth.
3. Every summary leaves things out
The answer to a misleading omission cannot be to include everything. A report that reproduces the whole archive has given the reader another archive to read.
What matters depends on the question. A history of the failed analysis should explain the original estimate and its withdrawal. A short report on the present evidence may omit the percentage altogether. A report about preservation of research files may properly concentrate on what was saved. Different purposes can justify different accounts of the same material.
Before assessing a report, we therefore need to know its question, audience, relevant date and scope. These should not be invented afterwards to make a disliked answer appear defective. The assignment itself can also be challenged: a narrowly framed question may leave out something the audience needs to know.
The proposed procedure adds a brief account record beside the prose. For our example, it could read:
- Question: Does the current evidence establish an improvement?
- Conclusion: The effect remains unresolved.
- Grounds: The review withdrew the estimate; the corrected comparison is unfinished.
- Qualification: Withdrawal of the estimate does not establish absence of a benefit.
- Revision: State that the estimate was withdrawn and the corrected effect remains unresolved.
The full proposal also records consequential relationships and reasons for excluding relevant material. It is intended to make the main conclusion open to examination, not to require a form for every sentence.
This record can be wrong too. A model might invent a conflict, mark a still-valid observation as superseded, or give a persuasive excuse for omitting contrary evidence. Its entries need sources and remain open to challenge. They describe proposed grounds for a report; they are not a trace of the model's internal computation.
4. Good editing is a serious alternative
Much of this sounds like ordinary good editing. It should. Careful reviewers already ask whether a conclusion is justified, whether qualifications survived the summary, and whether a correction goes too far.
Fact-checking need not stop at isolated sentences either. A sufficiently thorough check can identify an implied claim, state it explicitly and examine its grounds. The proposal does not depend on declaring existing review incapable of doing this.
The narrower question is whether making these relationships explicit in a maintained record helps enough to justify the work. A short checklist might do as well. So might a clear current-status conclusion with its supporting sources. If either is sufficient, a larger record may simply be unnecessary.
Nor should every objection trigger a rewrite. A reviewer may have noticed an irrelevant omission or a defensible difference of emphasis. The aim is an adequate answer to the question, not a report that gives equal space to every possible interpretation.
Corrections deserve the same scrutiny. In our example, replacing the suggestion of improvement with “the procedure does not work” would introduce a new unsupported conclusion. A useful review must examine what its own repair now invites the reader to believe.
5. A comparison that could reject the procedure
No experiment proposed in the paper has yet been run. It offers a design for finding out whether the account record helps, rather than evidence that it already does.
The first comparison gives every reviewer the same question and complete collection of source documents, called a dossier. Delivery must be checked: we cannot say a reviewer ignored a correction it never received.
The same model, with the same settings, reviews and revises reports in three ways: competent general review, review with a short checklist, and review with the account record. The ordinary reviewers may keep useful notes and correction links. The comparison must give the simpler approaches a fair chance to succeed.
The approaches receive comparable resource allowances. Making the record counts as part of its cost, as do model use, elapsed time and human work. An improvement bought with substantially more effort needs to be described as such.
Readers then receive the resulting reports alone. They do not receive the account record as extra evidence. This tests whether the procedure improves the report itself. Giving readers the record might also help, but that is a separate question, with an additional reading cost.
The study asks what readers understood and whether the dossier supports their answers. Separate assessors examine the reports themselves. The reference answers used to assess performance also need to be checked and challenged: human agreement does not settle what the evidence supports.
Fair, concise reports must be tested alongside defective ones. Otherwise a procedure that objects to every omission could look successful. The study must count mistaken objections, failed repairs, new errors and answers made too vague to be useful. Results from constructed examples and ordinary work should be reported separately.
Before the main test, the researchers must set thresholds for worthwhile improvement, acceptable harm and cost. If the results are precise enough to show no worthwhile benefit, the extra record should be dropped or redesigned. Equality between two poor methods would approve neither. If the evidence is too uncertain to distinguish them, the result is inconclusive.
6. Better models may need less help
Nothing in this proposal establishes that future AI systems will need the same procedure. A more capable model may preserve the relevant relationships through ordinary review. A simpler workflow may achieve the same result with less work.
That would be a reason to simplify. The research question concerns how adequately an account represents its evidence. The account record is one candidate response. We should not preserve it merely because we have given it a name or invested time in developing it.
Evidence of usefulness would also have a scope. A result with one model configuration and one kind of report would not establish value everywhere. The first study does not test retrieval quality, the passage of records between successive agents, or the benefit of using multiple models. Later work can examine those questions if the initial comparison warrants it.
There is a further danger: a record may display impressive care while leaving readers no better informed. Its explanations may themselves conceal omissions or unjustified connections. The proposal must remain subject to the criticism it directs at other reports. Documentation is useful only insofar as it serves the work being done.
7. Three related questions
CounterProof's broader proposed programme, Agentic Epistemics, asks how people and AI can reach warranted decisions and improve their work through criticism. Agentic Representation is one strand.
Agentic Stemmatics investigates transmission and relationships among texts and models, including what dependence means for the weight of agreement. Representation asks what an account presents its evidence as establishing. Agentic Institutions examines the arrangements through which records, corrections and working practices affect later decisions.
These questions interact in both directions. An institution selects what enters its library. A report can draw attention to a neglected objection. Evidence about shared origins can change how that report should describe apparent corroboration. Each programme still needs its own tests; success in one does not certify the others.
For Representation, the immediate task is modest: compare a specified way of reviewing accounts with competent, simpler alternatives. The result should help decide whether to use the procedure, simplify it or set it aside.
Reading further
Read the full research paper on this site →
The companion paper is Agentic Representation: Evidence, Selection and Warrant in AI-Assisted Reports, v0.3. Its DOI is 10.5281/zenodo.23727617. The paper was published on Zenodo on 28 October 2026; this introduction has no separate DOI.
For the distinctions behind the example, see §§1.1–1.3. For historical source criticism and the bounded use of Ankersmit, see §2.1. The proposed record and its limits are in §3; the comparison and stopping rules are in §§4.3–4.6. The distinction between the continuing question and the usefulness of this procedure is stated in §6.3.
Status and interests. This is an explanatory draft, not independent evidence for the paper. The study remains a proposal. Luuk Joseph Soons has a commercial interest in CounterProof's adversarial code-review practice. Prepared with Codex (OpenAI) at his request; responsibility for the argument and publication remains with the human author.