The Auditable Executable Paper

The paper breaks where the tools meet

A paper can contain a correct number and still be unable to explain itself.

Ask where Table 2 came from and the answer may cross a licensed extract, a private spreadsheet, a cleaning script, a notebook, a specialist solver, three manual classifications, a figure export, an email thread, and the manuscript. Each object may be locally correct. The scholarly chain can still be globally broken.

That break becomes visible when a source is corrected, a sample rule changes, a calibration is superseded, a reviewer challenges an assumption, or an author has to reconstruct an accepted result six months later. The team no longer follows one connected argument. It performs forensic file reconstruction.

AI increases the pressure. It can draft prose, formulas, code, tables, and interpretations faster than a research team can verify the relationships among them. A fluent paragraph may cite the wrong source. A plausible analysis may silently change a denominator. An apparently complete answer may convert missing or contradictory evidence into an ordinary value. More generated material does not repair a broken chain; it can make the break harder to see.

This is why Grids should not primarily be positioned as an AI that writes papers, an Excel replacement with more functions, or a substitute for every disciplinary tool.

Grids should not compete for the sentence, the cell, or the solver. It should compete for the connective tissue of scholarship.

The strongest category is the auditable executable paper: a citable, versioned research object in which the path from evidence to claim remains explicit, runnable, inspectable, and governed.

The promise is not automated truth. It is answerability.

A paper that can answer for itself

The words in the category matter.

• Auditable means a consequential number, figure, or claim retains its source, lineage, status, evidence boundary, review history, and authority. Auditability is broader than rerunning code.
• Executable means declared inputs, rules, and dependencies can recompute, or an external disciplinary tool can return a controlled execution receipt. It does not mean the paper acts autonomously.
• Paper means argument, authorship, citation, interpretation, and peer challenge remain central. Scholarship is not reduced to a pipeline.
• Living means a new source or assumption can produce a new state, mark dependent claims stale, and expose what changed. It does not mean a published result can be silently overwritten.
• Human-governed means computation can calculate, compare, validate, or propose. Qualified people and institutions retain authority over methods, interpretation, authorship, review, release, and consequential action.

A PDF, journal article, slide deck, repository deposit, or public web page becomes a release view of the research object. The accepted release stays immutable and citable. A correction or extension becomes a linked successor with its own evidence, not an edit that erases what prior readers saw.

Claim field	Example recorded value	Question the record can answer
Source identity	Dataset DOI, filing accession, archive object, or instrument run	Which exact evidence was used?
Method revision	Code commit, notebook environment, protocol, or interpretive rule	What transformed the evidence?
Execution receipt	Inputs, tool or runtime identity, output, warnings, and status	What actually ran, and what failed?
Claim boundary	Population, interval, uncertainty, sensitivity, and exclusions	What is the result allowed to mean?
Review state	Reviewer role, objection, response, approval, and remaining dissent	Who accepted which proposition on which record?
Release identity	Immutable version, date, contributors, and successor link	What did readers receive, and what changed later?

This distinction prevents two category errors. A live model is not automatically a publication, and a publication snapshot is not the whole research object.

The six-question test

The category can be tested against any important number, figure, or claim. If the object cannot answer these questions, the paper is not yet auditable and executable in the strong sense.

1. Where did this come from?

   The answer should identify the primary source or artifact, stable locator, source owner, observation or effective time, retrieved version, digest where appropriate, permissions, and every transformation between the source and the visible result.

2. What assumptions produced it?

   The answer should expose definitions, inclusion and exclusion rules, units, missing-data treatment, model specification, parameters, priors or weights, scenario choices, approximations, and the conditions under which the method should not be used.

3. What changed?

   The answer should show the earlier and later revisions, who or what supplied the change, why it was accepted, which dependencies recomputed, which claims became stale, and which released artifacts remain tied to the previous state.

4. What evidence contradicts it?

   The answer should retain counterevidence, incompatible definitions, failed runs, negative results, unresolved identity matches, reviewer objections, and uncertainty as first-class states. Contradiction should not be overwritten by the most recent or convenient value.

5. Who reviewed and authorized it?

   The answer should distinguish author, data steward, methods reviewer, domain reviewer, contributor, approver, and external authority. A recorded click is not meaningful approval unless the person had the role, evidence, and institutional authority to decide.

6. Can someone independently reproduce it?

   The answer should provide the data and source identifiers permitted for release, model or code revision, dependencies, runtime and environment, parameters, seeds, execution order, status, outputs, and comparison tolerances needed for an independent replay. When access, an ephemeral phenomenon, or a unique historical event prevents reproduction or replication, the limitation should be explicit.

The National Academies' 2019 consensus report on reproducibility and replicability is useful here because it distinguishes obtaining consistent computational results from the same inputs and methods from obtaining consistent findings through a new study. An executable paper may strengthen computational reproducibility while saying nothing by itself about empirical replication, causal identification, construct validity, or truth.

It does not make a claim true. It makes the claim answerable.

This category has a lineage

The auditable executable paper does not begin from a blank page. Jupyter notebooks combine narrative, code, and results. Quarto and related systems render computational documents. Repositories provide versioning and review around files. Research compendia package the materials behind an analysis. eLife and Stencila have demonstrated Executable Research Articles in which readers can inspect and rerun code and data behind figures.

Those are real advances, not foils.

They also show why execution alone is not the complete category. A 2019 large-scale study of Jupyter notebooks found that reproducibility depended on declared environments, data access, execution order, hidden state, and disciplined authoring practices. The lesson is not that notebooks fail. It is that a computational document still needs a complete contract around its inputs, dependencies, execution, and release.

The proposed Grid distinction is therefore narrower and more defensible than “a better notebook” or “a new paper writer.” It makes the connected research object the central abstraction: not only code beside prose, but source and assumption beside calculation; claim beside supporting and contradicting evidence; revision beside downstream consequence; review beside role and authority; release beside a complete reproduction receipt.

Other systems can implement parts of that abstraction, and some can be extended to implement much of it. The category claim is about what Grids should optimize for, not a claim that no prior tool can support provenance, reproducibility, review, or executable publishing.

What the surrounding tools remain good at

• Document editors remain the strongest general surface for readable argument, citation, tracked prose, and publication exchange. In the research object, the manuscript is a bound narrative view rather than a pasted endpoint.
• Spreadsheets remain excellent for accessible arithmetic, inspection, scenario work, and expert-owned models. They can be an input, a view, or an accepted reference implementation without becoming the entire provenance and review system.
• Notebooks remain authoritative environments for exploratory analysis, code, narrative, diagnostics, and many specialist methods. Their selected outputs and execution receipts can participate in the paper without being reimplemented.
• Repositories remain authoritative for source versioning, code review, issues, releases, and immutable commits. The research object adds semantic links from those artifacts to affected assumptions, outputs, and claims.
• AI writing and coding tools can propose text, formulas, mappings, tests, and alternative interpretations. Their outputs remain attributable proposals until sources, semantics, methods, and claims are reviewed.
• Disciplinary tools remain authoritative for validated representations and methods: statistical packages, GIS, qualitative coding systems, laboratory instruments, proof assistants, CAD and simulation systems, observatory pipelines, and HPC schedulers. Grids should connect to their selected artifacts and receipts, not pretend to replace their expertise.

Grids wins only if it makes continuity across those tools easier to inspect and govern than the file-and-copy process it replaces.

The research object needs an open skeleton

A credible category cannot depend on one vendor's private metadata. It needs portable identities and relationships.

The W3C PROV Ontology supplies a useful base vocabulary for entities, activities, agents, derivation, attribution, association, revision, plans, and primary sources. That is broader than a cell dependency graph. A formula can show which inputs it reads; provenance should also show which activity produced an artifact, which agent held which role, which plan governed the activity, and which revision superseded another.

The FAIR Guiding Principles add findability, access under explicit conditions, interoperability, and reuse supported by rich metadata, provenance, licenses, and community standards. FAIR does not mean every datum is public. Authentication, authorization, consent, rights, and restricted repositories still apply.

RO-Crate 1.3 describes a JSON-LD research-object package for distribution, reuse, publishing, preservation, and archiving. It can describe files and linked resources, the people and organizations involved, equipment and software, funding, citations, licenses, and reuse conditions. DataCite's metadata schema supports consistent identification, citation, and retrieval. CRediT provides structured contributor roles. CITATION.cff makes software and dataset citation human- and machine-readable. JATS remains an important interchange format for journal content.

These standards are not badges to place on a marketing page. They are category requirements and exit paths:

• Can the research object export its provenance without flattening roles and activities into comments?
• Can another repository identify, cite, preserve, and relate an exact release?
• Can a journal receive a valid publication snapshot that resolves back to the accepted executable revision?
• Can contributors review their role assertions?
• Can restricted data remain governed while useful metadata and qualified relations remain available?
• Can a capable third party reconstruct the object without an undocumented vendor intervention?

The current Grid proposition should be evaluated against those questions. The existence of a reactive model, history view, or signed package does not establish PROV-O, RO-Crate, DataCite, CRediT, CFF, or JATS conformance.

An editorial stress test across eighty-one fields

The University of Texas at Austin's 2026–27 graduate catalog lists 81 majors that offer a PhD. We used that public list as a coverage frame for an internal editorial exercise, drafting one bounded design hypothesis per field.

The working notes are not published or independently reviewed, so they are not evidence that Grid supports 81 fields. The exercise helped us look for a cross-disciplinary invariant and select contrasting examples for this paper. The examples below remain design propositions for evaluation, not deployments, validated workflows, or measured outcomes.

The methods changed dramatically across fields. The invariant did not:

Every important scholarly claim should retain a navigable relationship to the sources, assumptions, computations or interpretive steps, external artifacts, revisions, disagreements, and responsible human judgments that produced it.

Three examples show why the category is broader than a computational notebook.

Accounting: from filing accession to table cell

An empirical accounting paper may begin with SEC filings and licensed datasets, then cross SQL, hand-coded classifications, sample filters, derived variables, estimation software, robustness scripts, and manuscript tables. The SEC's EDGAR APIs expose submission history and XBRL facts, but an API response is only the beginning of the scholarly chain.

In an auditable executable paper, a reported coefficient or descriptive cell can resolve to the filing accession, taxonomy and fact context, amendment state, retrieved source digest, sample rule, variable definition, estimation specification, software receipt, output, and accepted manuscript revision. Every exclusion is a named rule rather than an undocumented row deletion. Reported, restated, standardized, and researcher-adjusted values remain distinguishable.

If an amended filing changes a constructed variable, the paper does not silently refresh Table 2. It creates a new source revision, shows the affected sample and estimates, marks dependent claims for review, and preserves the release that used the earlier filing. Independent statistical software remains the benchmark for the selected method. Accountants and domain researchers—not the runtime—decide construct validity, interpretation, and professional meaning.

History: from archival object to explicit inference

A historical argument is executable in a different sense. It may connect archival objects, people, places, uncertain date intervals, identity decisions, transcriptions, translations, claims, and counterevidence. The task is not to turn interpretation into a formula. It is to keep the path from source to inference inspectable without collapsing uncertainty or disagreement.

IIIF Presentation 3.0 provides a standardized way to present compound digital objects across institutions. The W3C Web Annotation Data Model can attach an annotation to a resource or selected segment. PREMIS 3.0 provides preservation metadata around objects, events, rights, and agents.

Within the research object, a chapter claim can resolve either to an archival object and exact region or to an explicitly labeled scholarly inference. Conflicting accounts remain separate. An uncertain date remains an interval. A curator-approved identity split can invalidate part of a chronology and mark dependent narrative claims stale. Archivists and community custodians retain access and description authority; historians retain responsibility for identity, significance, and interpretation. AI transcription or entity suggestions remain provisional contributions, never source testimony.

Computer science: from run environment to claim gate

A systems paper may separate source code, build environment, workloads, benchmark data, CI or HPC jobs, seeds, hardware, raw logs, summary scripts, figures, failed runs, and claims. A successful rerun of one notebook does not establish that the experiment matrix is comparable or complete.

The auditable executable paper treats the experiment as a contract: hypothesis, system variants, workload matrix, environment manifest, metrics, seeds, time budget, failure policy, acceptance rules, and claim links. Existing CI and HPC systems remain the execution authority. The proposed implementation would receive version-pinned artifacts and typed summaries, explicitly retain failed and negative runs, recompute comparisons, and bind every figure point to a run, environment, and raw log. Those retention and figure-lineage requirements are not implied by ordinary model evaluation and must be implemented and tested.

The ACM artifact review and badging policy provides a useful external review frame. The paper should be able to support an independent stratified rerun without presenting a machine-checkable claim gate as peer review. Authors and reviewers approve benchmarks, threat models, generated code, and claims. Operators authorize execution capabilities. An AI assistant cannot self-certify correctness or security.

The same invariant travels further

• Economics: the AEA data and code policies make precise documentation and nonexclusive access to data and code an explicit publication concern. A Grid design can connect source vintage, sample waterfall, identifying assumptions, estimates, robustness results, and counterfactuals while leaving specialist estimators authoritative.
• Operations research: a formulation-to-certificate registry can retain immutable instances, generators, solver settings, seeds, bounds, feasibility checks, timeouts, and hardware metadata around benchmark practice such as MIPLIB. A solver result remains conditional on the represented objective and constraints.
• Astronomy: FITS files, IVOA provenance, calibration state, reduction pipelines, catalog releases, selection functions, fits, and figures can remain connected while raw data and high-volume processing stay in observatory systems.
• Mathematics: MathML 4 expressions, definitions, assumptions, conjectures, examples, counterexamples, exact symbolic records, numerical experiments, and external Lean proof artifacts can share a dependency graph without relabeling computational evidence as proof.
• Chemistry: an IUPAC digital-standards spine can connect reagent identity, lot and purity, reaction steps, samples, instruments, calibration, processing, characterization, and claims. The model can flag an unsafe or unsupported condition; it cannot authorize laboratory work or declare compound identity.
• Public policy: facts, legal authorities, fiscal models, effect estimates, uncertainty, implementation constraints, distributional effects, nonmonetizable impacts, stakeholder values, and public comments can remain separate rather than disappearing into one score. Accountable public officials retain policy authority.

The breadth matters because “executable” cannot mean “Python runs.” In some fields it means recalculation. In others it means a proof artifact verifies, a chronology invalidates, a source contradiction remains live, a policy rule replays by effective date, or a specialist system returns a signed result.

Execution is not truth

The article's strongest safeguard should also be its clearest product boundary.

• Dependency lineage explains the represented model. It does not establish causal truth or prove that important variables were included.
• Reproduction can faithfully reproduce a flawed method, biased sample, invalid instrument, or mistaken interpretation.
• Compiler and runtime validation establish technical conformance to a declared contract. They do not establish scientific validity, fairness, ethics, or policy legitimacy.
• A digital signature can establish identity and byte integrity. It does not establish safety, peer review, preservation, or scholarly merit.
• An optimization result is optimal only within the represented objective, constraints, tolerances, solver status, and search bounds.
• A reporting checklist can expose missing items. Completing it does not validate a study.
• A human-in-the-loop button is ceremonial unless the reviewer has sufficient evidence, a real role, and actual authority.
• A shared model does not authorize centralizing protected, licensed, export-controlled, culturally sensitive, or community-governed data.

An auditable executable paper is therefore not a truth machine. It is a disciplined way to show what was represented, what ran, what failed, what changed, what remains disputed, what was omitted, and who was responsible for the accepted claim.

What is documented now—and what the category still has to earn

This category proposal separates three layers that should remain visible in every academic proposition.

Documented Grid foundations

Current first-party Grid documentation describes a bounded set of relevant primitives: reactive models with named dependencies and typed values, saved history, dependency explanation, and validation, evidence-aware predicates, controlled external functions, multiple surfaces bound to one model, and version-pinned decision-package contracts. Those documented primitives make the category plausible.

The documentation does not yet describe a complete scholarly receipt, the standards mappings below, or an auditable-executable-paper product. Those are proposed requirements, not shipped capabilities.

They are product evidence, not independent outcome evidence. They do not show that a university has adopted the workflow, that a paper became reproducible, or that a researcher saved time.

Academic adaptation required

A credible implementation still needs a complete, versioned scholarly receipt and tested mappings or exports for provenance, research-object packaging, persistent identity, contributors, software citation, journal interchange, preservation, rights, and discipline formats. That includes, where appropriate, PROV-O, RO-Crate, DataCite, ORCID, CRediT, CodeMeta, CITATION.cff, JATS, preservation profiles, and domain-specific adapters.

It also needs explicit scholarly review states, permissions, anonymized or open-review patterns, correction and retraction behavior, records retention, portable export, and independent replay across supported deployment targets.

Until those pathways exist and pass round-trip tests, the category is a roadmap with documented foundations—not a finished academic platform.

Human and institutional governance

Method validity, peer review, authorship, IRB and consent, privacy, accessibility, academic integrity, Indigenous and community authority, professional judgment, licensing, security, records obligations, and publication acceptance remain with qualified people and institutions.

Grids should represent those boundaries and retain their evidence. It should not claim to replace them.

Start with one completed paper

The category is ambitious, but the first evaluation should be deliberately small.

Choose one completed study with frozen inputs, an accepted reference implementation, and no live participant, grading, publication, or policy consequence. Select one or two consequential findings—not the entire research program.

Before configuring Grid:

1. Freeze the accepted manuscript, source and data manifests, code and environment, specialist-tool outputs, and expected numbers.
2. Name every material definition, exclusion, assumption, parameter, uncertainty expression, reviewer role, and authority boundary.
3. Record the current baseline: time to reconstruct the result, undocumented manual steps, inconsistent denominators, broken links, review cycles, and corrections that cannot be traced to affected outputs.
4. Decide which disciplinary systems remain authoritative and which selected inputs, outputs, manifests, or receipts Grid may use.
5. Define independent comparison tolerances and kill criteria before seeing the Grid result.

The compact gate below keeps a polished rerun from standing in for a complete evaluation.

Gate	Passing evidence	Automatic failure
Claim identity	Every selected claim resolves to an immutable release, claim class, scope, and responsible contributors	A claim exists only as prose copied from an unversioned draft
Source and method	Exact source, selection, exclusion, transformation, method, environment, and parameter revisions are recoverable	A consequential input or manual step has no identity or owner
Change reach	A seeded correction marks every dependent value, figure, claim, review, and release state affected	A released surface remains silently bound to a superseded result
Review and authority	Objections, dispositions, approval scope, remaining dissent, and exact accepted revision are attributable	Access credentials or an AI proposal are treated as scholarly approval
Replay and export	An independent evaluator reproduces within declared tolerances and inspects a portable research-object package	Reproduction depends on undocumented vendor intervention or hidden state

Then test the six-question contract. Seed a corrected source, changed exclusion rule, stale calibration, missing dependency, fabricated citation, failed run, contradictory source, unresolved reviewer objection, restricted field exposed to the wrong view, and AI-written claim presented as approved.

Require an independent reviewer to:

• reconstruct every selected result from the accepted release;
• identify the source and assumptions behind each important value;
• locate every downstream consequence of a controlled change;
• see negative, failed, stale, and contradictory evidence without opening a forensic collection of files;
• distinguish calculation, recommendation, interpretation, review, and authorization;
• verify that protected or licensed material stayed inside its approved boundary.

Scale only if the selected results independently reproduce within declared tolerances, every released value has a complete lineage or an explicit missing state, every seeded contradiction remains visible, access tests pass, and reviewers can explain a changed claim without undocumented intervention from the vendor.

A passing fixture supports a bounded product-demonstration claim for that paper and environment. It does not establish scientific validity, independent replication, institutional compliance, usability at scale, instructional effectiveness, research productivity, or improved scholarly outcomes.

The category shift

AI can help write a sentence. Spreadsheets can calculate. Notebooks can execute code. Repositories can version files. Disciplinary systems can perform the specialized work. Documents can carry the argument into review and publication.

The unsolved coordination problem is keeping those objects connected strongly enough that the paper can answer for itself. Trace a Research Claim from Source to Reviewed Release follows one synthetic correction from evidence through an immutable successor release. Make the Structure Visible describes the underlying model discipline, while How to Evaluate Executable Decision Infrastructure turns the proposition into a bounded test. Teams can take the circulation-ready white paper into review, record the fixture in the decision evaluation worksheet, and compare the authority model in The Authority Contract for AI-Assisted Decisions.

That is the category Grids should pursue:

The PDF is a release. The auditable executable paper is the research object.

Do not ask only whether AI can write the paper. Ask whether every important number, figure, and claim can explain where it came from, what assumptions produced it, what changed, what contradicts it, who authorized it, and how another scholar can independently reconstruct it.

Inspectability and repeatability

Make the work behind the finding inspectable. The companion research brief connects specific methodological concerns to Grid’s academic and research features. It includes cited research, worked examples of uncertainty and experiment comparisons, and an explanation of what preserving a result allows another researcher to inspect.

Repeatability is central to the argument: the brief explains the conditions for replaying eligible saved computations and distinguishes that from independent reproduction and empirical replication. It identifies Grid Pro research workflows and proposed extensions separately.

Download the research inspectability and repeatability brief (PDF, 9 pages).

References

1. National Academies of Sciences, Engineering, and Medicine, Reproducibility and Replicability in Science, 2019.
2. World Wide Web Consortium, PROV-O: The PROV Ontology.
3. GO FAIR Foundation, The FAIR Guiding Principles.
4. RO-Crate Community, Research Object Crate 1.3.
5. DataCite, DataCite Metadata Schema.
6. NISO, CRediT Contributor Role Taxonomy.
7. Citation File Format Community, Citation File Format.
8. NISO, Journal Article Tag Suite.
9. eLife and Stencila, Welcome to a New ERA of Reproducible Publishing, 2020.
10. Pimentel, Murta, Braganholo, and Freire, “A Large-scale Study about Quality and Reproducibility of Jupyter Notebooks”, 2019.
11. U.S. Securities and Exchange Commission, EDGAR Application Programming Interfaces.
12. Association for Computing Machinery, Artifact Review and Badging.
13. International Image Interoperability Framework, Presentation API 3.0.
14. World Wide Web Consortium, Web Annotation Data Model.
15. Library of Congress, PREMIS Data Dictionary for Preservation Metadata, Version 3.0.
16. American Economic Association, Data and Code Policies and Guidance.
17. IAU FITS Working Group and NASA, FITS Standard.
18. International Virtual Observatory Alliance, Provenance Data Model.
19. World Wide Web Consortium, Mathematical Markup Language 4.
20. Lean project, Lean documentation.
21. IUPAC, Digital Standards.
22. University of Texas at Austin, Graduate Degrees Offered and Degree Programs, 2026–27.
23. MIPLIB, Benchmark Set.
24. Grid Developers, Product concepts.
25. Grid Developers, Explain and validate a decision.
26. Grid Developers, Predicates.
27. Grid Developers, External functions.
28. Grid Developers, Surfaces.
29. Grid Developers, Decision packages.