The Minimum Viable Scientific Record

Operationalizing 30 years of consensus.

Summary

The open science movement does not lack a shared vision. The 153 documents analyzed here were selected from a collection of over 2,000 public texts about open science.[1] The corpus shows substantial agreement across institutions and constituencies. It agrees that openness is a means to better science rather than an end in itself.[2][3][4][5] It agrees that the research record should become a linked, machine-actionable web connecting the full range of research outputs, the work that produced them, and their subsequent assessment and use.[5][6][7] It agrees that research assessment is a binding constraint on implementation.[2][3][4][5] And it broadly agrees that shared infrastructure should be community-governed and protected against private enclosure, although the sources propose different arrangements for ownership, financing, and commercial participation.[2][3][5][8][9][10]

Decades of declarations, statements of principle, and calls to action have failed to put that vision into practice. Declarations recommend reforms without making anyone’s funding depend on them, research assessment still rewards publication in prestigious journals, and many services that openness relies on depend on short-term funding. Where shared records did become routine, for example in the Human Genome Project and the Protein Data Bank, funders and journals tied sharing to funding or publication, applied the requirement across the field, paid for the means to comply, made compliance visible, and gave researchers a say in the rules.[11][12] The results of these shared records in the two examples were faster discovery, more reproducible research, worldwide collaboration, new medicines, and, decades later, AlphaFold, which learned to predict protein structures from the Protein Data Bank.[11][12][13]

This essay proposes a minimum viable scientific record (MVR) that would bring that kind of structured record to every field of research, so researchers and AI systems could build on, train on, and validate it at a greatly accelerated pace. This MVR would make the open science vision possible, though putting the vision into practice would require funders make operating on it a condition of their money.

The MVR is the smallest version of the linked, machine-actionable record the open science movement has described. It connects each claim to the work and evidence behind it, names who and what did the work, records who checked it and what they found, states who may reuse it, and remains available across services. The existing record grew one journal, repository, and database at a time; the MVR would be designed as a single system of interoperable components, built to a shared specification. Early versions of its components already exist. In this essay’s own analysis, for example, several components were used to extract 21,422 claims from the 153 analyzed documents, and the published record links each claim to its source passage, documents the methods, and lets readers rerun the counts.[1] Building the MVR may mean extending and connecting the rich existing ecosystem of such components rather than starting from scratch.

The rapid development of AI makes operationalizing this record urgent, both for checking the research AI produces and for giving AI systems dependable materials to work and train from. AI systems can now produce plausible papers, analyses, and reviews far faster than people can check them.[14][15] Checking will increasingly fall to software, including AI systems, and software can check only what it can read. The MVR records what was done, by whom, with which data and methods, and what has already been checked, in a form both people and software can use.

The same record would also make each part of the open science vision possible. Anyone could verify a result and validate or challenge the work. Outputs could be released as soon as they exist, reused under stated terms, and shared and checked without a publication fee; every contribution would be attributed; and copies held by many services would keep any single operator from controlling access.

While the technology necessary for the MVR can be built, and perhaps already exists, making the record routine depends on decisions that funders control. They should pay for a shared specification of the record and for the tools and services researchers need to contribute to it. They should require every grant to add its work to that record and to make its outputs openly available without delay, at no cost to researchers. Each grant should name who must add the work, by when, how compliance will be checked, and how the result will affect payment, renewal, or eligibility. Because the record attributes every contribution, funders should make institutional awards depend on hiring and promotion that use it to assess all the roles research depends on, from data and software to reviewing and replication. And they should help the shared services that hold the record become financially sustainable while keeping them under community governance.

None of this requires new principles. The movement has agreed on those for decades; encapsulating them behind different words “for the age of AI” will change nothing. Putting them into practice requires funders to decide that their next grant agreements will pay for the record, require it, and check it.

1. What the corpus already agrees on

The analysis covers 153 public documents, from declarations and funder policies to consultation responses and infrastructure principles. Model-assisted comparison grouped statements judged equivalent across at least two documents and organized the groups into themes. Under each heading, the evidence gives the number of documents and claim groups in the theme and names well-known documents among them, and the list that follows gives the claims that the most documents share. A document may support only part of a heading’s position. The methods explain the counting rules, how the lists were chosen, and the limitations.[1]

In short, the headings state what the corpus agrees on, and the lists show the claims behind each position. Readers who already accept the headings can skip to section 2, which covers the choices that agreement leaves unresolved, or, if they know those debates, to section 3.

Openness serves better science

90 of 153 analyzed documents make related claims · 209 claim groups[1]

Major documents: UNESCO, BOAI20, Vienna Principles, Barcelona Declaration, ORFG

The corpus values openness as a means to better research, because it lets others inspect, test, reuse, correct, and extend the work.[2][3][4][5][6] Its most widely shared claims hold that openness:

The scientific record should become linked and machine-actionable

60 of 153 analyzed documents make related claims · 78 claim groups[1]

Major documents: FAIR Principles, FORCE11, Barcelona Declaration, UNESCO

The corpus calls for a record in which claims, data, software, and workflows can be identified, cited, and linked, and for open information connecting outputs to the people, organizations, and funding behind them.[5][6] Its most widely shared claims hold that:

Research should be available without delay

56 of 153 analyzed documents make related access claims, including “as soon as possible” · 45 claim groups[1]

Major documents: SPARC, OSTP Nelson Memo, COAPI, Gates Foundation, UNESCO

The corpus asks for research, including its data and code, to be available immediately, so that others can begin scrutiny and reuse without waiting.[16][17] The sources set different release points, from publication or acceptance to “as soon as possible.”[17][18][19] The corpus’s most widely shared claims hold that:

Research should be open to reuse

55 of 153 analyzed documents make related claims · 62 claim groups[1]

Major documents: SPARC, Plan S, Science Europe, UNESCO, COAPI

The corpus asks for open licenses that permit reuse and for machine-readable, nonproprietary formats that make reuse technically possible, because free reading access alone establishes neither.[16][17][19] Its most widely shared claims hold that:

Sharing research should not depend on the author’s ability to pay

31 of 153 analyzed documents make related claims · 32 claim groups[1]

Major documents: SPARC, Plan S, BOAI, Gates Foundation, Berlin Declaration

Publication fees can keep researchers who cannot pay from contributing work that others could use.[2][3] The corpus asks for a route that costs authors nothing, such as depositing in a repository or posting a preprint.[16][18][20][21] Its most widely shared claims hold that:

Research should remain identifiable and connected across systems

45 of 153 analyzed documents make related claims · 63 claim groups[1]

Major documents: FAIR Principles, Barcelona Declaration, ORFG, UNESCO, SPARC

The corpus asks that outputs, people, organizations, and grants carry persistent identifiers and open metadata, so that different services can refer to the same object and link information about it.[3][5][6][19] Its most widely shared claims hold that:

Research assessment should recognize the full contribution to science

54 of 153 analyzed documents make related claims · 150 claim groups[1]

Major documents: DORA, The Metric Tide, LERU, UNESCO, BOAI20

The corpus asks assessment to recognize the work that others rely on, including data, software, protocols, and reviewing, and to judge its quality rather than the reputation of the journal where it appears.[2][3][4][22] Its most widely shared claims hold that:

Shared infrastructure should remain accountable to the research community

46 of 153 analyzed documents make related claims · 135 claim groups[1]

Major documents: POSI, IOI, Barcelona Declaration, UNESCO, SCOSS

The corpus asks that shared services keep the record available through changes in their funding, leadership, or ownership, and that they remain accountable to the communities that rely on them.[2][3][5][8] Its most widely shared claims hold that:

2. Where agreement leaves choices unresolved

Agreement that research should be accessible and reusable leaves open how services are organized, which publishing costs institutions should pay, who may authorize reuse, how common standards accommodate different communities, and what role commercial organizations may play.[2][3][9][10][18][23][24] Under each heading, the list gives the most widely shared claims on that choice, each with the number of documents that make it.

Shared infrastructure and distributed control

25 of 153 analyzed documents make related claims · 21 claim groups[1]

Major documents: POSI, Bengaluru Roadmap, FORCE11, BOAI20

The corpus wants infrastructure that no single operator controls and that is not fragmented into services that cannot work together. It leaves open which functions to share, who controls them, and how services divide responsibility.[2][3] Different documents in the corpus hold that:

Which publishing costs institutions should pay

43 of 153 analyzed documents make related claims · 73 claim groups[1]

Major documents: BOAI20, SPARC, GAO, Plan S, Bengaluru Roadmap

The corpus agrees that article-processing charges (APCs) exclude researchers and institutions with less money. It divides over what to do with the money now spent on publication, including whether to keep paying hybrid journals, subscription journals that charge a fee to open individual articles.[1] Different documents in the corpus hold that:

Who may authorize access and reuse

43 of 153 analyzed documents make related claims · 97 claim groups[1]

Major documents: CARE, Barcelona Declaration, ORFG, POSI

The corpus accepts that some data must stay closed but divides over who decides. Some documents rely on legal and privacy safeguards; others give that authority to the communities the data come from, including Indigenous peoples.[3][5][19][23][25] Different documents in the corpus hold that:

Common standards and different research contexts

44 of 153 analyzed documents make related claims · 51 claim groups[1]

Major documents: UNESCO, OCSDNet, POSI, CARE

The corpus wants shared standards that let services exchange information but rejects a single correct way to practice open science. The choice is who sets common standards and how they accommodate different disciplines, regions, and communities.[2][3][5][6][9][24][26] Different documents in the corpus hold that:

Commercial participation and community ownership

13 of 153 analyzed documents make related claims · 21 claim groups[1]

Major documents: UNESCO, Bengaluru Roadmap, BOAI20, Toluca–Cape Town Declaration

The corpus favors community-owned, non-commercial infrastructure but differs on how far commercial providers may take part. The choice is what private actors may own or control, which services may charge, and what rights the research community keeps if an operator changes.[3][9][10] Different documents in the corpus hold that:

3. Why agreement has not made open practice routine

Open practice has not become routine because institutions have not turned their commitments into requirements they enforce, contributions they reward, and services they sustain. Researchers cannot make those institutional changes through individual willingness to participate.

Declarations leave implementation to later decisions

The major declarations describe reforms without making anyone’s funding or eligibility depend on carrying them out. BOAI20 recommends actions to other organizations, the Vienna Principles describe how scholarly communication should work, and UNESCO encourages Member States to pursue reforms.[2][3][4] Each leaves institutions to decide how those expectations will affect their own rules and resources.

Signatories of the Barcelona Declaration go further and commit to specific actions, but the declaration specifies no penalty for failing to carry them out.[5] An institution can therefore commit in public without changing the conditions attached to its funding, procurement, or research assessment.

Research assessment still gives researchers conflicting incentives

Researchers can support openness while receiving little professional recognition for practicing it. In the 2025 State of Open Data survey, with more than 4,700 responses from 151 countries, 81 percent of respondents supported open data, while 69 percent believed researchers received too little credit for sharing it.[27]

eLife’s experience shows the cost of changing publication practice while evaluation remains tied to journals. In 2023 it began publishing reviewed preprints with public reviews and assessments, and it later became ineligible for a Journal Impact Factor following changes to its indexing status.[28][29] Submissions fell 55 percent between 2024 and 2025. In eLife’s assessment, institutional evaluation had not kept pace with public commitments to reform and still depended on journal-level metrics in some settings.[28]

Infrastructure depends on continuing financial commitments

The services that make sharing possible need support that outlasts individual projects, because a repository must stay available, preserve what it holds, and keep the information that lets later researchers find and use it. Public support can be withdrawn. The Dutch ministry cut Open Science NL’s annual allocation from €20 million to €10 million beginning in 2025, and the Dutch Research Council largely offset the cut, bringing the program’s 2026–2027 budget to €36 million.[30] A grant condition requiring researchers to share their work must therefore come with support for the services entrusted with preserving it.

4. How institutions made shared records routine

Shared records did become routine in the Human Genome Project and the Protein Data Bank, where funders and journals made sharing a condition of decisions they already controlled and paid for the means to comply.

The Human Genome Project made data release a funding condition

Funders used grant conditions to require sequencing centers to share data while the Human Genome Project was underway. In 1997, the National Human Genome Research Institute required specified DNA sequence data to be released publicly within twenty-four hours. The Wellcome Trust incorporated the data-release principles agreed at a 1996 meeting in Bermuda into its grants.[11]

NIH and Wellcome funded nearly 90 percent of the project’s human-genome sequencing effort, so these conditions reached most of the publicly funded work.[11] Because the main public funders applied the same release rules, a publicly funded center that shared its data was not competing against public centers that withheld theirs. From 1997, a center’s allocation of further sequencing work depended on the data it had released during the preceding year.[11]

The public funders also financed the sequencing centers, databases, and operating support needed to release and use the data. Participating researchers debated and voted on the release rules, could comment before their publication, and revised them at subsequent meetings.[11]

Journals made the Protein Data Bank part of publication

Most journals publishing new macromolecular structures require authors to deposit the underlying structural data, experimental data, and metadata in the Protein Data Bank. Publication depends on a deposit the journal can check. Some journals also require the archive’s validation report with the manuscript.[12]

Funders supported the professional staff who checked submissions, maintained the archive, and updated older records. Regional centers used common processing standards so that researchers and software could use data deposited by different laboratories.[12]

Researchers help set the archive’s requirements. Proposed changes to its coordinate format are published for community comment, and advisory committees are consulted before implementation.[12]

Institutions paired obligations with the means to comply

In both cases, institutions:

  1. tied the obligation to a decision they already controlled, the allocation of further sequencing work or the publication of an article;
  2. applied it across the field, through the main public funders’ shared release rules and most journals’ deposit requirements;
  3. paid for the means to comply, including sequencing centers, databases, and the staff who check and maintain the archive;
  4. made compliance observable, through public releases, deposits that journals can check, and validation reports;
  5. gave researchers a role in setting the rules, through votes on the release rules and public comment on format changes.[11][12]

Funders seeking a minimum viable scientific record should take the same five steps. They must also judge the requirement itself. A condition can be enforced without producing useful science, and compliance alone does not establish that a rule is fair to the communities it affects.

5. Why AI makes the minimum viable scientific record more urgent

AI systems are beginning to produce research, and they already work from the research record and learn from it. The open science vision called for a record that people and machines could inspect, reuse, and contest.[6] AI makes that record necessary, both for checking the research AI produces and for giving AI systems dependable materials to work from.

AI produces research faster than people can check it

Sakana’s experimental AI Scientist was supplied with an existing codebase and used literature search, code modification, and computational experiments to produce research reports.[31] Generative AI can also produce plausible papers, analyses, and reviews at low cost, whether or not the research behind them is sound.[14]

Lovén argues that the cost of producing convincing scientific text is falling without a comparable reduction in the cost of checking it.[14] The Research on Research Institute’s 2026 report warns that reviewing each output separately cannot absorb the expected volume of research.[15] People will therefore need software, including AI systems, to share the work of checking. Software can check only what it can read, so the data, code, methods, versions, and previous checks behind a result must be available in a form it can use.

The MVR provides that form. It lets people and software trace a finding to the work that produced it and determine what a claimed check actually established. When an AI system produces a result, the record shows which materials it used and what it did with them, so others can verify and validate its work as they would a person’s. Its output can then become input to further research, so a missing provenance record, an unsupported claim, or an unrecorded correction affects both that output and the work that relies on it.

AI systems work from the record

Where relationships among claims, inputs, and versions remain implicit, a system may have to infer them, and an inferred relationship can be wrong. Explicit links would let it retrieve the materials associated with a claim instead.

AI systems learn from the record

The research record is also training material for AI systems. When data and their relationships are available only through the text, tables, or figures in a PDF, training teams must extract and reconstruct them before using them as structured inputs. Extraction cannot recover materials that were never reported or shared. AlphaFold, which predicts protein structures, was trained on structural data accumulated in the Protein Data Bank, including the atomic coordinates that describe protein structures.[13] That leap was possible because, for decades, journals had required authors to deposit their structures and funders had paid for the archive that checked and maintained them.[12]

The MVR would bring that kind of structured record to every field, extending it beyond data to epistemics, validation, attribution, and process. Training teams could use that record to assemble examples that associate a reported finding with its supporting materials instead of reconstructing those relationships from prose.

6. A minimum viable scientific record

The minimum viable scientific record is the smallest version of the linked, machine-actionable record the open science movement has long described.[5][6][7][32] It gives a persistent identifier to everything the research used and produced, including data, code, methods, samples, and publications, in every version, and to the contributors and grants behind it. It connects each claim to the prior work it builds on, to the exact inputs and outputs behind it, and to what follows, including reviews, verifications, validations, corrections, disputes, and reuse, stating how each is related. It names the people, AI systems, and institutions responsible for the work and attributes each contribution. It shows how the work was produced and records each check, including who or what performed it, how, what they found, and what has not yet been checked. It states who may reuse the materials, on what terms, and whose permission is required, and it identifies restricted materials even when their contents remain closed. People and software can retrieve the whole record, move it between services, and keep it available if a service holding it closes. With it, anyone can redo the work, see whether a claim holds up, credit those who did it, and build on it. The record makes the work checkable; it does not certify that the work is correct.

This essay proposes six requirements for the MVR, each of which a funder can check.

  1. Every input and output, each version, each contributor, each organization, and each grant has a persistent identifier that resolves.[5][6][19]

  2. Every part of the record links to the parts it depends on and to what follows it. Claims link to the prior work and evidence behind them, outputs link to the inputs and versions that produced them, and each links to the reviews, verifications, validations, corrections, disputes, and reuse that follow.[7] Each link states the relationship, such as whether evidence supports or opposes a claim, whether a dataset was used, or whether a review challenges a result.[6][32]

  3. Each part states its license, attribution, consent, and access conditions in a form software can read. Restricted materials name the authority that grants permission and how to request it.[23]

  4. The record shows how the work was produced and changed, including its source materials, significant transformations, and previous versions. It names who or what performed each step and who is responsible for it.[6][32]

  5. Each check records who or what performed it, the version examined, the method, date, result, and limitations. Expected checks not yet performed are listed. Corrections and disputes stay attached to the check they concern.[15][33]

  6. Software can retrieve authorized parts of the record through open, documented interfaces, one item at a time or in bulk. An export keeps everything above intact, so another service can take it over.[5][8]

The existing scientific record grew one journal, repository, and database at a time. The MVR is designed as a single system of interoperable components. Each field can contribute its own repositories, tools, and standards, as long as they conform to the shared specification (section 8.1).

7. What the MVR makes possible

Each position in section 1 needs something the record provides. The record gives people and software, including AI systems, the materials behind a result and the history of every previous check and assessment. Each new check or judgment is added to the same record, so later users can inspect the evidence, repeat a check, or challenge a judgment.[6][15]

Verify results

Verification examines whether a scientific output exists, and whether a reported result follows from the declared data and methods. A person or a computational service, including an AI system, can run the specified code on the recorded data and compare its output with the reported value. The check may reproduce that value, produce a different result, or fail because required information is missing. The same check applies whether a person or an AI system produced the work.

Validate and challenge the work

Validation examines whether the methods suit the question and whether the conclusions hold, including through independent replication. Reviewers can use the identified materials and recorded checks when assessing data quality, methods, interpretation, and ethical concerns.[15] Each assessment records the reviewers, their criteria, reasoning, and relevant conflicts.

Different communities can apply different criteria to the same work. The record keeps each assessment, including disagreements, together with its grounds. Corrections and disputes stay attached to the result they concern, so later users can see how its standing changed.

An organization could also certify that an identified version of the work met a stated standard, based on specified checks and reviews. The certification would make a defined claim about the work rather than guarantee that everything was done correctly.

Release research without delay

Because every version of an output has its own identifier and links (section 6), a dataset, piece of code, or result can enter the record as soon as it exists instead of waiting for an article about it. The record also shows which checks each version has received and which are still expected, so early release does not present unchecked work as checked. The Human Genome Project’s sequencing centers released specified data within twenty-four hours, under grant conditions from funders who also paid for the databases that received them.[11] Whether outputs are released without delay depends on the conditions funders set (section 8.2).

Reuse and build on the work

The record helps researchers find the materials behind a finding, determine which versions they need, and identify the conditions governing reuse. Data, software, methods, and reviews can be cited and connected to later work, preserving attribution to the people who produced and maintained them.[5][6][19]

Reuse must also preserve the authority and obligations attached to the materials. CARE’s authors warn that machine learning accelerates data reuse in systems that can disregard relationships, unequal power, and the circumstances in which data were collected.[23] The Bengaluru Roadmap similarly warns that AI systems can absorb open scholarship into closed commercial platforms without adequate attribution, transparency, or reciprocity.[9] The record attaches consent, community authority, and conditions of use, including any required consultation or return of benefits, to the materials and to any recorded copies, transformations, or models trained on them. Where tracking is lawful and safe, it can also record later uses. The record makes these obligations visible; the people and institutions reusing the material remain responsible for honoring them.

Share research without charging authors

Repository deposit and preprints already let researchers share work without paying a publication fee. Review and recognition, however, still depend heavily on journal publication in many settings. In the MVR, reviews, checks, and attribution attach to the record itself, so work shared without a fee can be reviewed, verified, and attributed without passing through a journal. Hosting, checking, and review then no longer have to be bought together from one publisher. Many parties can provide them, including research communities, repositories, independent checking services, and reviewers whose contributions the record attributes.

Recognize the full range of contributions

The record identifies contributions to data, software, methods, curation, reviewing, replication, correction, and validation, attributes each to the people who made it, and connects it to the work that relied on it.[5][6][19] Institutions could use the record to assess that work alongside conventional publications, judging its quality directly instead of relying on publication counts or journal prestige.[2][3][5] Their assessment policies would still determine how that work is evaluated and whether it receives credit.

Keep infrastructure accountable to the research community

The implemented record can be exported intact and taken over by another service, and nothing in it requires a single operator. Copies can be held by many services and institutions at once, so no single service can cut off access to it. BOAI20 warns that centralization can create lock-in, and the Bengaluru Roadmap calls for decentralized, mirrored architectures to keep infrastructure resilient and sustainable.[2][9] If an operator raises prices, changes its terms, or closes, the communities that rely on it can move their part of the record, including its links and checking history, to a service they govern. Holding copies in many places does not by itself govern the record or keep its services running.

Accelerate research

The record would speed research by removing work that researchers now repeat. The corpus counts faster research among the benefits of openness (section 1). Shared data alone does not deliver that. In one journal issue, only one of fourteen articles with open-data badges was exactly computationally reproducible, and three more were essentially reproducible with minor deviations.[34] Anyone building on the others would first have to reconstruct how the work was done. The MVR records the code, versions, and steps that reproduction needs, along with the checks already performed, so researchers can build on a result as soon as it is released instead of reconstructing it.

The same structure speeds building on earlier work. AlphaFold was trained on decades of structured deposits in the Protein Data Bank (section 5). A record that links results to their data, methods, and checks would let researchers and AI systems train on, combine, and extend earlier work in every field in the same way.

Checking and correction would move faster too. Software can verify a result as soon as its data and code enter the record, and a correction or dispute is visible from every recorded study that used that result.

8. How funders can operationalize the vision

The implemented MVR makes possible each position the open science movement agrees on. Whether those positions become routine practice depends on decisions that funders and institutions control. Funders would need to pay for a shared specification and the tools that meet it, require each grant to add its work to the MVR, make institutional funding depend on assessment that uses the record, and help the infrastructure that holds it become financially sustainable.

8.1 Fund a coordinated program to make the MVR operational

Funders should coordinate a development and integration program that makes the requirements in section 6 practical in ordinary research. The Bengaluru Roadmap calls for shared infrastructure to be developed as “a coordinated, interoperable ecosystem rather than a collection of fragmented or duplicative systems.”[9] Development funding for research infrastructure is usually awarded to one service at a time, and each improvement is proposed and reviewed on its own merits without having to fit a common design. Many existing services were built to deposit and cite files, before anyone specified a record like the MVR. Improving them separately adds formats and interfaces that someone must later connect, and can make a working system harder to assemble.

The program should therefore begin with a shared specification, developed and governed with the research communities that will use it, that turns the six requirements into technical standards, common interfaces, and conformance tests. A specification in this sense is not a statement of principles. It is a detailed document, often hundreds of pages, that states what each part of the record must do and why, how the parts work together, and which tests show whether an implementation conforms.

The specification should cover only what every part of the record must carry and exchange, leaving each field’s own standards in place. It should define what the record must do, not which systems provide it, so that existing services and new architectures are judged by the same tests. Funders should pay to adapt existing services where they can conform, fund new components where they cannot, and make further development funding for any service depend on conformance. The specification requires no single operator or platform. Operating support for the services that preserve deposited materials continues under section 8.4.

Tools should capture identifiers, grant associations, versions, and other routine information automatically, and let researchers inspect and correct the links between claims and supporting materials. Information entered once should serve deposit, publication, grant reporting, and assessment.

The program should also connect the verification and validation services described in section 7. Anyone verifying or validating the work, whether a person, a service, or an AI system, should be able to retrieve the materials they need and add their results to the record, linked to the exact version they examined. Restricted materials remain subject to their stated access conditions. Because these services can affect funding and publication decisions, their methods, error rates, and corrections should be open to independent review, and affected researchers should be able to appeal. Funders should pay for mandatory checks in ways that protect the checkers’ independence.

Work recorded with one combination of conforming tools should remain usable and preservable through another, with its identifiers, relationships, provenance, permissions, and checking history intact.

Funders should pay researchers, maintainers, and affected communities to take part in designing, testing, and governing the program’s tools and services, and should check whether cost, language, accessibility, or connectivity excludes anyone. Program funding should cover development, integration, training, and operation during implementation; the arrangements in section 8.4 should cover continuing costs. Independent evaluators should test the services in real projects across disciplines, and funders should expand them only when the evidence shows they work, can be sustained, and are practical for the people expected to use them.

8.2 Make open access and contribution to the MVR conditions of each grant

Every grant should require recipients to add each material research output to the MVR, maintain their part of the record, and make those outputs openly available. The obligation concerns the completeness and accuracy of the record, including honest reporting of failed or unperformed checks. Any requirement for further assessment or certification should be specified and funded separately.

Each grant agreement should name who must add the work to the record, set delivery dates, cover the costs, define how compliance will be checked, and state how payment, renewal, or eligibility depends on the result. Funders should adopt these conditions jointly with other funders, so that researchers who comply are not competing for positions and grants against researchers who do not.

Every output in the record should be openly available without delay, through rights retention or another lawful mechanism, unless a recorded restriction applies. Default licenses should be CC BY for written work; CC0 for data and metadata where law, ethics, third-party rights, and legitimate community authority permit; and an Open Source Initiative–approved license for code where the recipient has the right to grant one. Restrictions should be recorded with a named authority and an access process. Recipients who build models or services from shared materials should identify their sources and honor the attached conditions. Where community agreements require benefits to be returned, such as improvements, annotations, or corrections, funded programs should pay for that work.

Compliance should cost researchers nothing. Some funders already separate compliance from publication fees. The Gates Foundation expects funded manuscripts to be shared as preprints and does not pay APCs, and Astera excludes work on journal articles from its funds and staff time.[35][36]

Automated checks should flag missing deposits and other defects, and independent auditors should test a sample of records for accuracy. Recipients should have time and support to fix ordinary problems. Consequences should be proportionate to the seriousness and persistence of any remaining failure and may include withholding final payment or denying renewal. Recipients should not be penalized for failures caused by services they were required to use. Funders should publish the findings of these compliance checks and audits.

8.3 Make assessment reform a condition of institutional funding

Funders cannot require open records while allowing universities to penalize the researchers who create them. Institutions seeking center-scale awards, training grants, infrastructure awards, or other institutional funding should show that their hiring, promotion, and assessment use the record to recognize work on data, software, methods, curation, reviewing, replication, correction, and validation alongside publications.

Signing a commitment such as the San Francisco Declaration on Research Assessment (DORA) is not enough. Institutions should provide their current policies, reviewer instructions, and assessment templates, and explain how committees judge the quality of contributions without relying on journal prestige or output counts.

Funders should begin with a pilot of three to five universities with different missions, resources, and national evaluation settings, and pay for the work of revising policies, training committees, and adapting information systems. An independent team should compare written policies with a confidential sample of actual decisions, including how committees evaluated contributions to the scientific record. McKiernan and colleagues provide a precedent for examining promotion and tenure documents for reliance on Journal Impact Factors.[37]

An institution whose practice does not match its commitments should receive time and support to correct it. If material discrepancies persist, it should lose eligibility for the designated awards until it can show the change. Funders should hold their own review panels to the same standard.

8.4 Establish durable funding for the infrastructure

The services that hold the record need income that outlasts any single grant. The Principles of Open Scholarly Infrastructure recommend sustainable income for ordinary operations and time-limited funds only for time-limited activities.[8] Funders should use infrastructure awards to help services reach that position, starting with services that meet or are adopting the specification in section 8.1 and can show who relies on them and what would be lost if they stopped.

Where a service’s income is insecure, a transition award should pay for the work of establishing it, such as recruiting institutional members, negotiating service agreements, or securing continuing public support. Funders should share that work by bringing together the institutions that depend on a service and coordinating multi-year commitments. The Global Sustainability Coalition for Open Science Services (SCOSS) offers a precedent. It requests an initial three-year commitment, sets contribution bands for institutions with different resources, and states that it aims to develop arrangements that sustain services beyond the initial support period.[38] Awards should set milestones and review whether commitments and income actually materialize.

UNESCO identifies direct funding and earmarking a share of grants as possible mechanisms for supporting open science infrastructure.[3] For initial budgeting, this essay proposes 1 to 2 percent of annual research-grant value, checked against the documented costs of the services grants require and coordinated with other funders of the same services.

Funded services must keep their public function. Operators should gain no exclusive control over publicly funded metadata, should operate under accountable community governance and transparent ownership, and should keep access, export, and transfer obligations through any change of ownership or policy, subject to lawful restrictions and community authority.[8][9] If an operator fails, funders should pay to transfer the record and essential services to another. Some essential public functions may need continuing public or philanthropic support; funders should agree in advance how that support is renewed and reviewed, so that operation does not depend on repeatedly proposing new projects.

9. The decision before funders

Open science has spent decades arguing that research should be available for others to examine, reuse, and challenge. Endorsing that ambition does not by itself change what researchers must produce, how their contributions are assessed, or how the services they depend on are funded.

The minimum viable scientific record would let others trace a project’s claims, examine its history, check it, and build on the work. Requiring that record would give institutions an observable basis for judging whether their commitments to openness have been implemented. As AI systems begin to produce and use research, the same record would let people and software check what those systems did and would give those systems dependable materials to work from.

Requiring researchers to contribute to the record also gives funders a responsibility to make compliance feasible. They should finance a shared specification and the tools and labor that contributing requires, require every grant to add its work to the record with immediate, openly licensed access unless a recorded restriction applies, make institutional funding depend on assessment that recognizes the people doing that work, and help the services that preserve the record become financially sustainable. The next grant agreement should make these responsibilities explicit. It should fund support and set delivery dates at the outset, and tie payment or renewal to evidence that the work has been added to the record.

Contact

jon@scios.tech

Disclosure

The author is co-director of SciOS, which publishes this essay. The author is also the instantiating steward of the Institute of Open Science Practices (IOSP) and a contributor to the Modular Interoperable Research Attribution (MIRA) schema. The essay cites IOSP and MIRA.[39][32] The claim graph published with this essay uses the Resilient Research Graph Infrastructure (RRGI) record format, which SciOS develops. This work received no funding.

Sources

  1. Author’s 153-document corpus, collection inventory of 2,098 collected texts, and methods. The analysis contains 21,422 extracted claims. The theme counts, claim groups, and claim graph provide the supporting records. The complete package contains every file and the commands to rerun the counts.
  2. Budapest Open Access Initiative, 20th Anniversary Recommendations (15 March 2022). https://www.budapestopenaccessinitiative.org/boai20/
  3. UNESCO, Recommendation on Open Science (adopted 23 November 2021, 41st General Conference). Official legal text. Development and adoption history. Implementation and four-year reporting process. Cited provisions include paragraphs 8, 9, 13(b), 14(f), 17, 18, 18(d), 18(e), 20, and 23–24.
  4. Open Access Network Austria, The Vienna Principles (2016). https://viennaprinciples.org/. https://doi.org/10.5281/zenodo.55597
  5. Barcelona Declaration on Open Research Information (2024). Commitments at https://barcelona-declaration.org/commitments/. Background at https://barcelona-declaration.org/background_and_context/. Zenodo record at https://doi.org/10.5281/zenodo.10958522.
  6. FORCE11, “Improving the Future of Research Communication and e-Scholarship” (2011). https://force11.org/info/force11-manifesto/. https://doi.org/10.4230/DagMan.1.1.41
  7. Octopus, a primary research record in eight linked publication types. https://www.octopus.ac/ and https://www.jisc.ac.uk/innovation/projects/octopus-creating-a-new-primary-research-record-for-science
  8. Principles of Open Scholarly Infrastructure. Version 2.0 (October 2025). https://openscholarlyinfrastructure.org/. https://doi.org/10.14454/G8WV-VM65. Version 1.0 at https://openscholarlyinfrastructure.org/posi-v1.0/. Adopters and self-audits at https://openscholarlyinfrastructure.org/adopters/.
  9. Bengaluru Roadmap and Action Plan on Diamond Open Access (2026). https://doi.org/10.5281/zenodo.20168763. https://scienceeurope.org/media/0wkapl1e/bengaluru-roadmap-and-action-plan-on-doa.pdf
  10. Toluca–Cape Town Declaration on Diamond Open Access (December 2024), with participant commentary through SPARC (2025). https://sparcopen.org/news/2025/toluca-cape-town-declaration-is-a-milestone-for-recognizing-science-as-a-public-good/
  11. Maxson Jones, Ankeny, and Cook-Deegan, “The Bermuda Triangle: The Pragmatics, Policies, and Principles for Data Sharing in the History of the Human Genome Project,” Journal of the History of Biology 51:693–805 (2018). https://doi.org/10.1007/s10739-018-9538-7. https://pmc.ncbi.nlm.nih.gov/articles/PMC7307446/. The “Measured Consensus” section, pp. 750–751, documents participant debate, the vote, objections, and the opportunity to comment on the final statement before release. Official implementation history appears in NHGRI’s “Reaffirmation and Extension of NHGRI Rapid Data Release Policies”. The 1997 Second International Strategy Meeting report, pp. 1–2, records participant-endorsed quality, error-resolution, and sequence-allocation rules.
  12. Burley et al., “Protein Data Bank: the single global archive for 3D macromolecular structure data,” Nucleic Acids Research 47:D520–D528 (2019). https://doi.org/10.1093/nar/gky949. See also RCSB PDB history, funding and institutional information, and Burley et al., “RCSB Protein Data Bank: Sustaining a Living Digital Data Resource That Enables Breakthroughs in Scientific Research and Biomedical Education”, Protein Science 27:316–330 (2018). For governance, the worldwide Protein Data Bank publishes proposed format changes for community comment, accepts comments for sixty days, and consults its Advisory Committee and partner scientific advisory committees before implementation.
  13. Jumper et al., “Highly Accurate Protein Structure Prediction with AlphaFold,” Nature 596:583–589 (2021). https://doi.org/10.1038/s41586-021-03819-2
  14. Lovén, AI-Augmented Science and the New Institutional Scarcities, arXiv:2605.02566v3 (revised 19 July 2026). https://arxiv.org/abs/2605.02566v3
  15. Pinfield, Waltman, Kolarz, Gregory, and Brasil, The Future of Peer Review: A System-Wide Perspective (Research on Research Institute, June 2026). https://doi.org/10.6084/m9.figshare.32685351
  16. Coalition of Open Access Policy Institutions, response to the 2020 OSTP RFI (22 April 2020). https://sparcopen.org/coapi/activities/coapi-response-to-2020-rfi-public-access-to-scholarly-information/
  17. SPARC, “Comments in Response to OSTP RFI 85 FR 9488” (6 May 2020). https://sparcopen.org/wp-content/uploads/2020/05/SPARC-OSTP-Public-Access-Response.pdf
  18. SPARC, response to NIH RFI NOT-OD-25-138 on publishing-cost caps (15 September 2025). https://sparcopen.org/wp-content/uploads/2025/09/SPARC-NIH-RFI-Response-APC-Price-Caps-9.2025-1.pdf
  19. Open Research Funders Group, response to the OSTP Scientific Integrity Policy Framework RFI (8 April 2022). Official response page. See also ORFG’s community call on research-output tracking and community response.
  20. SPARC agency-response series, 2023–2025, collected at https://sparcopen.org/our-work/2022-updated-ostp-policy-guidance/
  21. ASAPbio, response to the NIH Plan to Enhance Public Access (30 March 2023). https://asapbio.org/asapbios-response-to-the-nih-plan-to-enhance-public-access/
  22. National Academies Roundtable on Aligning Incentives for Open Science, Guide to Supporting Open Scholarship for University Presidents and Provosts (2021). https://www.heliosopen.org/s/Open-Scholarship-Support-Guide-v6721.pdf
  23. Carroll, Herczog, Hudson, Russell, and Stall, “Operationalizing the CARE and FAIR Principles for Indigenous data futures,” Scientific Data 8:108 (2021). https://doi.org/10.1038/s41597-021-00892-0
  24. OCSDNet, Open Science Manifesto (2017), based on consultations across 26 countries. https://ocsdnet.org/manifesto/open-science-manifesto/
  25. CARE Principles for Indigenous Data Governance. Carroll et al., Data Science Journal 19:43 (2020). https://doi.org/10.5334/dsj-2020-043. See the Global Indigenous Data Alliance’s current CARE Principles page and the Research Data Alliance’s preserved one-page summary.
  26. Albornoz, Okune, and Chan, “Can Open Scholarly Practices Redress Epistemic Injustice?” in Reassembling Scholarly Communications (MIT Press, 2020). https://doi.org/10.5281/zenodo.4320060
  27. Digital Science, Figshare, and Springer Nature, The State of Open Data 2025, tenth edition. https://doi.org/10.6084/m9.figshare.30823079. The official release reports more than 4,700 responses across 151 countries and the cited percentages.
  28. Rodgers and Pattinson, “The eLife Model: Three-year update,” Inside eLife (17 March 2026). https://elifesciences.org/inside-elife/7da92a51/the-elife-model-three-year-update. The journal’s indexing-status notice describes the Clarivate decision.
  29. San Francisco Declaration on Research Assessment, “Clarivate’s Actions Regarding eLife: DORA’s Response” (25 November 2024). https://sfdora.org/2024/11/25/clarivates-actions-regarding-elife-doras-response/
  30. Open Science NL, Work Programme 2024–2025. https://zenodo.org/records/10074873. Work Programme 2026–2027. https://zenodo.org/records/17579089. The Dutch government’s 2025 budget memorandum records the reduction.
  31. Lu et al., The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery, arXiv:2408.06292v3 (revised 1 September 2024). Primary paper.
  32. Modular Interoperable Research Attribution, or MIRA, project site and preliminary schema snapshot at commit 483f0b21480c0f4ced9c1519fc0f6df2b8617cfe (retrieved 21 August 2026).
  33. National Information Standards Organization, “Establishing Content Trust Markers” (April 2026), an active standards-development project. https://www.niso.org/niso-io/2026/04/establishing-content-trust-markers
  34. Crüwell, Apthorp, Baker et al., “What’s in a Badge? A Computational Reproducibility Investigation of the Open Data Badge Policy in One Issue of Psychological Science,” Psychological Science 34(4):512–522 (2023). https://doi.org/10.1177/09567976221140828
  35. Gates Foundation, 2025 Open Access Policy. Compliance guide and publishing-fee guidance.
  36. Astera Institute, Open Science Policy (10 December 2025). https://astera.org/open-science-policy/. https://doi.org/10.5281/zenodo.17873235
  37. McKiernan, Schimanski, Muñoz Nieves, Matthias, Niles, and Alperin, “Meta-Research: Use of the Journal Impact Factor in academic review, promotion, and tenure evaluations,” eLife 8:e47338 (2019). https://doi.org/10.7554/eLife.47338
  38. SCOSS, “Key Facts and Figures”, “How It Works”, and contribution bands (retrieved 21 August 2026). SCOSS reports pledged amounts and states that each infrastructure collects funds directly.
  39. Institute of Open Science Practices, “A Theory of Change for Science” (interactive narrative, 2026), especially the account of plural scientific systems connected through a shared technical layer. https://www.iosp.science/theory