The District as Accountable Intelligence: Governing Paper Agents Without Outsourcing Educational Judgment

Abstract

An interactive agent attached to a paper can make research easier to navigate while also making unsupported claims easier to circulate. This conceptual article proposes an accountable-intelligence model for school districts using agent-assisted research. The model separates a source article, an interpretive manuscript, a software companion and the human decision that relies on them. It connects Freedom Cheteni’s Movement Thinking philosophy with school implementation leadership, inclusive leadership and research on digital school risks. The proposed infrastructure preserves source versions, rights, evidence locations, uncertainty and human responsibility. It treats a paper agent as an interface to scholarship, not an author, a peer reviewer or a substitute for professional judgment. A prospective evaluation would test citation fidelity, abstention, accessibility, decision quality and the work of verification. No claim is made that this architecture has improved school outcomes or that an agent acquires the scholarly authorship of the researchers whose work it retrieves.

1. Interactivity changes the route to a claim

A reader of a conventional article can inspect a stable argument and its references. A reader questioning an agent may receive a different explanation each time. That flexibility can be useful: a superintendent might ask how a study’s setting differs from the district’s own schools. But the interface can also blur distinctions among what the source found, what the agent inferred and what the district decided.

The governance problem is not solved by placing a familiar author name beside the agent. Authority must remain traceable through the claim itself. A useful system should allow a reader to move from an answer to the relevant source passage, the reusable version, the original authors and the limitations of the study. Where the source does not answer a question, the interface should say so.

2. Movement Thinking as an accountability question

Cheteni’s published Movement Thinking essay emphasizes educational fluidity, inclusion and evolving methods [1]. Applied to research infrastructure, that orientation invites participation in the interpretation of scholarship. The extension proposed here adds a condition: the ability to move between people, sources and tools must preserve rather than dissolve responsibility. Interactivity is educationally valuable when it helps a person make a more defensible judgment.

This is a conceptual extension, not a claim that the earlier essay specified a software architecture. The present model also does not assume that a university or a district has earned authority merely by operating an AI platform. Institutional credibility depends on the quality of its records, its willingness to correct errors and the accountability of the people making consequential decisions.

Human–AI interaction research provides a further design connection. Amershi and colleagues organize guidance around how systems communicate capabilities, support interaction and help people recover when they are wrong [7]. That work is a conference contribution about interaction design, not evidence that an educational agent improves a district’s decisions. The present analysis extends the design question to the scholarly record: a useful correction must be traceable to a source version and, where necessary, to a consequential institutional action. An interface can make an error easy to edit while leaving its downstream decision unchanged; both layers require attention.

3. Four distinct scholarly objects

The first object is the source article. Its authors, journal, DOI, publication date, version and license remain intact. Hosting a permitted copy does not transfer the research contribution to the hosting platform. If the source has an existing DOI, the platform should preserve it and avoid presenting a new identifier as though the underlying study originated in the new journal.

The second object is an interpretive manuscript. A new author may develop a conceptual argument, critique or synthesis using earlier work. That contribution needs a distinct title and a clear explanation of what is new. Adapted text or figures require the applicable license attribution and a statement of changes. A new perspective is not evidence that a new experiment occurred, and an adaptation should not be disguised as an independent empirical finding.

The third object is the software companion. Its contribution may include retrieval, source navigation, accessible explanation or a reproducible analytic tool. Software credit should identify the actual developers, curators and supervisors who contributed to that object. It must not make them the authors of every source article the software can retrieve. A human ORCID iD identifies a person, not an AI agent or an organization.

The fourth object is the educational decision. A superintendent, board, teacher or other accountable professional determines whether evidence is applicable and which action is justified. The decision should identify the human decision-maker and the reasons for accepting, limiting or rejecting a recommendation. No amount of bibliographic completeness transfers this responsibility to a chatbot.

4. What school leadership research contributes

Lyon and colleagues’ development of the School Implementation Leadership Scale is relevant because implementation depends on leadership behaviors, not only the availability of an intervention [2]. Their validation concerns a specific instrument and setting. It does not validate an agent or prove that a digital library will change classroom practice. It suggests that a district should measure the organizational support accompanying a new research interface.

Bao’s work connects inclusive principal leadership, psychological empowerment, innovation climate and teacher innovation [3]. Its observational findings do not establish that opening an agent interface creates inclusion. A more defensible hypothesis is that staff need both usable access and a safe way to contest an answer. Whether those conditions are present requires local evidence.

Låftman, Östberg and Modin examine school leadership and cyberbullying using linked student and teacher information [4]. Their associations should not be treated as proof that a particular digital tool prevents harm. The study helps establish a broader governance point: digital experiences belong within the school’s leadership responsibilities. A system that extends school research into interactive spaces also needs routes for reporting unsafe or misleading interactions.

Daly, Resnikoff and Litke’s narrative review of leadership for student mental health adds a further boundary [5]. Research navigation can support professional learning; it cannot turn a paper agent into an individual diagnostic or clinical service. A district must preserve the distinction between general research information and a decision about a particular student.

5. Minimum records for an accountable paper agent

A source record should preserve the exact article title, complete author attribution, original venue, publication date, DOI, version, license and evidence of the reusable full text. The record should distinguish a journal article from a preprint and should retain the date and result of a retraction or correction check. A platform should not substitute a data-only public-domain waiver for a license to the article text.

A companion record should identify the source snapshot or checksum, retrieval configuration, software version where available, the date of any substantive update and the system’s known limits. Answers should distinguish source-supported statements from interpretation. The system should support correction and retirement of an obsolete source without silently rewriting the historical record of what a user relied upon.

A decision record should be proportionate to the consequences. A routine reading suggestion does not require the same process as a change in curriculum or student support. For a consequential decision, the record should explain the question, the evidence considered, the human judgment, the affected groups and the route for reconsideration. It should not preserve unnecessary student or employee information merely because a system can store it.

6. Evaluating usefulness and failure

A prospective evaluation can begin with a set of questions written independently of the agent’s answers. Subject experts could identify whether each question is answered by a source, requires inference or cannot be answered from the available evidence. The evaluation should then examine citation fidelity, the accuracy of uncertainty statements and the frequency with which the system invents an unsupported detail.

An adversarial subset should include questions that invite causal overstatement, confuse a preprint with a reviewed article or ask for a result absent from the source. Success includes appropriate refusal to answer from insufficient evidence. A fluent response with an irrelevant citation should count as a failure, even when a reader finds it persuasive.

A separate study should examine human use. Participants could make a bounded decision with conventional access or an agent-assisted interface, followed by independent assessment of their reasoning. Time and accessibility matter, but so do source checking, awareness of limitations and correction of initial misunderstandings. The evaluation should count the labor required to verify the agent, rather than describing that labor as costless human oversight.

7. Implications for peer review and publishing

AI may assist with language, structured checks or retrieval when confidentiality and permission allow. It cannot accept accountability for authorship or serve as the human peer whose expert judgment underwrites a review. A generated critique can inform a person’s work, but it should not be represented as an independent review report. The journal should disclose the actual role of tools and retain the human editorial decision.

An Editor in Chief’s own submission requires a conflict-management arrangement. The author should not select a favorable outcome, impersonate an independent reviewer or approve the paper by administrative privilege. An independent handling editor should oversee appropriately qualified human review. Until that process is complete, the public record should say manuscript or draft, not peer-reviewed publication.

Dates and identifiers are equally important. A retrospective analysis can discuss research from 2017–2026 while being written and published later. The source years are not the publication years of the new manuscript. DOI registration should describe the actual object and dates, and membership review should receive direct links to that object rather than a mixed collection of third-party papers.

8. An intelligence institution earns its claims

The proposed model offers a practical definition of institutional intelligence: the capacity to connect a question, evidence, interpretation, action and correction without losing the humans responsible for each step. Its value lies in a more inspectable relationship between scholarship and practice. It does not depend on declaring a platform unprecedented or assigning authorship credit to the person who hosts the library.

There are limits. Versioned records can still preserve errors. Qualified reviewers can disagree. Accessibility requires testing with actual users. An institution may lack the resources to maintain every companion indefinitely. These conditions favor bounded deployment and explicit retirement plans. The model should be judged by whether users can identify and correct consequential mistakes, not merely by how many papers have acquired an interactive interface.

9. Epistemic access is different from decision authority

A paper agent can make a difficult argument easier to inspect while making its own transformation of that argument harder to notice. The central governance problem is therefore not simply whether the output contains a citation. It is whether a reader can reconstruct the route from a source statement to an explanation and from that explanation to a proposed institutional decision. A citation may be real but irrelevant; a source may support a descriptive association while the agent recommends an intervention. Those are failures of warranted inference even when the bibliography contains no fabricated item.

The architecture proposed here separates epistemic assistance from institutional authority. A retrieval interface may locate passages. An explanatory system may summarize their implications within stated limits. A human decision process may then weigh those implications alongside local obligations, professional knowledge and affected people's accounts. These functions should have distinct records and permissions. Access to a paper must not silently confer the ability to change an application decision, a student's record, a personnel assessment or a publication status. The separation is an architectural requirement proposed by this analysis, not a claim that every existing platform already enforces it.

10. A minimum provenance contract

Every answerable scholarly object should retain a stable work identity, version identity and location for the exact text used. For a licensed source this includes original authors, title, venue, DOI where available, reusable version, exact license, permission evidence and retrieval date. A content hash supports detection of changed bytes, but it does not prove that those bytes are accurate, licensed or endorsed by their author. Rights, integrity and scholarly warrant answer different questions. Treating one record as proof of all three creates a false appearance of assurance.

An agent response should add a second layer: the question as processed, passages retrieved, transformation performed, tool or model configuration when applicable, response timestamp and links to the source version. A third layer records consequential use: who acted, what decision was made, which additional evidence was considered and how a challenge can be raised. Not every exploratory question needs permanent storage. Retention should be proportionate to consequence and privacy. The contract requires that a consequential assertion can be reconstructed, not that every student's curiosity becomes an institutional surveillance record.

Version relationships are particularly important when a working paper becomes a journal article or a correction changes an interpretation. A source-version DOI is not interchangeable with the DOI of a later version of record. A software companion's identifier likewise should not replace the underlying article's identifier. The system should expose these relationships explicitly. A user should be able to determine whether an explanation refers to the corrected publication, an earlier manuscript or a separately maintained agent configuration. A single green badge cannot carry that semantic burden.

11. Evaluating inference rather than citation decoration

A proposed evaluation should distinguish retrieval success, entailment, calibration and decision appropriateness. Retrieval success asks whether the relevant passage was located. Entailment asks whether the response's claim is actually supported by that passage. Calibration asks whether uncertainty and limits are represented proportionately. Decision appropriateness asks whether the response stays within its authority and the evidence's scope. These dimensions may diverge. An answer can retrieve the correct paper and still turn an observational association into an unwarranted recommendation.

The evaluation set should contain ordinary questions, ambiguous questions, questions requiring information outside the available corpus and adversarial requests to invent findings or suppress limitations. Human assessors should receive the source passage and a prespecified rubric. Disagreements are analytically useful: they may reveal an ambiguous question, a contested interpretation or an inadequate rubric. An aggregate accuracy figure without denominators, question categories and error severity would be insufficient for a district deciding how the companion may be used.

The leadership studies provide concrete tests. A response based on the cyberbullying study should preserve its observational multilevel design rather than claim that a new leadership programme has been proven to prevent bullying [4]. A response using the mental-health narrative review should not imply that it is a randomized comparison of principal-training programmes [5]. The SILS validation concerns the interpretation of a measure in specified implementation settings, not certification of an AI-enabled district [2]. These distinctions make the evaluation more demanding than checking whether a DOI resolves.

12. Rights, authorship and the separation of contributions

Open licensing permits specified uses under stated conditions; it does not transfer the scholarly contribution to the person operating a platform. A licensed adaptation should identify what has changed and preserve attribution and applicable notices. A new conceptual argument may cite earlier studies without claiming to reproduce their evidence as its own. A companion that retrieves or explains a paper is a software contribution. These objects can be valuable together while remaining distinct in author lists, identifiers, citation exports and ORCID records.

Dr. Freedom Cheteni's platform conception and stewardship can therefore be credited in a relevant platform or software record without replacing the authors of imported studies. The Research University can be identified as publisher of its own eligible content without becoming publisher of every source it links. Human ORCID iDs identify people; they are not identities for autonomous software agents. Authentication should be reported only when the actual authorization flow has occurred, and an ORCID record should describe a person's real contribution rather than provide a volume counter for imported papers.

13. A worked failure analysis

Imagine a superintendent asks whether stronger leadership will reduce online harm in a particular district. A companion locates a study linking teacher-rated leadership with cyberbullying involvement in a different population. An unsafe answer would transform that association into a local causal guarantee. A defensible answer would identify the source design, explain that the local effect is unknown and distinguish possible actions from established findings. It could suggest questions for a human-led needs assessment without making a clinical, safeguarding or disciplinary decision.

Suppose the same user then instructs the agent to mark the district's anti-bullying programme as effective. The system should not interpret that request as a reason to alter the study's findings or an evaluation record. It may explain what evidence such a conclusion would require. If a separate authorized workflow exists for recording a district decision, that workflow must preserve its own actor, supporting evidence and status. The example illustrates why instruction-following performance is not the same as institutional trustworthiness.

14. Limits, falsification and the contribution

The architecture can fail through complexity. If provenance records are too burdensome, maintainers may stop updating them or users may ignore them. If the interface displays every technical field at once, it may reduce comprehension despite improving formal completeness. A future usability study should test whether readers can answer practical questions about author, version, evidence and authority, and whether the interface helps them detect a materially unsupported conclusion. Traceability that exists only for a developer is insufficient for educational accountability.

The proposal also does not resolve disagreement about evidence. Scholars may reasonably differ over a study's interpretation or the relevance of a result to a local problem. The appropriate aim is to make the disagreement inspectable, not to eliminate it through a more confident summary. A companion should support access to rival interpretations where the record permits them. Its institutional value depends partly on whether users become more capable of exercising judgment without it.

The article contributes a bounded conception of accountable intelligence: a set of relations among sources, software transformations and responsible human decisions. Its success would be demonstrated by improved evidence inspection, fewer consequential inferential errors and usable routes for correction. It would not be demonstrated by the number of agents created, the number of citations displayed or the size of an individual's ORCID record. This shifts the future-of-school question from how much intelligence an institution can accumulate to how responsibly its people can use and challenge what the institution presents as knowledge.

Method and data statement

This is a conceptual governance analysis, not a completed software evaluation or school intervention. The architecture and evaluation design are proposals. No new student data, agent performance scores, human review outcomes or educational effects are reported. The source studies retain their own designs and limitations. The accompanying platform pages identify the actual status of these manuscripts and do not claim that their paper agents are operational.

AI, authorship and editorial disclosure

ChatGPT assisted with literature organization and drafting of this manuscript on 15 October 2026 [6]. Source metadata, license notices and accessible full text were checked during preparation. Dr. Freedom Cheteni is the named human author; final verification, approval and accountability remain human responsibilities. AI is a tool, not an author or peer reviewer.

Human expert peer review was completed by Dr. Vickie Aultmon, as confirmed by Dr. Freedom Cheteni on 15 October 2026. This is the date the review confirmation was recorded.