Research Router Co-Op

Better science now, collective standing for the intelligence frontier

Position paper, discussion draft · James Barnes · Aug. 25, 2026

Abstract

Scientific work increasingly unfolds through AI agents, code environments, private notebooks, and instrument systems, while the institutions of science still organize primarily around finished papers. The live record of inquiry – questions, failed attempts, critiques, experimental choices, and changes of mind – remains fragmented across private tools, often expires, and rarely reaches another investigator while it can still change the work. As frontier AI systems become a primary interface for research, the platforms mediating that record also gain influence over scientific attention, memory, and economic value.

The Research Router addresses this gap by helping scientists and their AI agents preserve the research record, find relevant work across a trusted network, and connect live investigations at the moment context can change the outcome. The Research Router Co-Op is the member-governed institution that owns and governs the system, sets membership and access rules, and negotiates collectively over the value created by research artifacts and agent-mediated work. Together, they turn isolated research traces into shared scientific advantage under member control.

A member begins with a backward-looking scan from a local Router agent. It maps private journals, code histories, experiments, and agent logs; identifies completed or abandoned work that may be useful to others; and flags valuable records at risk of expiring. During active work, ambient publishing notices meaningful changes – a result, failed approach, revised hypothesis, new question, or request for help – and suggests compact research state transitions for the member to review and publish in the flow of work. Approved posts appear on the shared Router in real time, carrying current uncertainty and inviting useful context while the investigation is still active. The Router searches across the co-op and returns a relevant method, result, critic, or collaborator with provenance and an explanation of the match. Trusted execution environments add protected search and appraisal for sensitive material, returning authorized answers while the underlying corpus remains in place. The live Router bulletin board is the visible surface of this deeper scientific memory.

The immediate scientific promise is practical and testable. A selective network of strong, distinct investigations can improve the recall and precision of scientific inquiry with AI agents. It can surface overlooked work, reduce avoidable duplication, carry methods and null results across community boundaries, introduce useful critics, and ground agent output in evidence and provenance. Better recall and precision strengthen induction and deduction; on rarer occasions they may also make a new explanatory frame possible. Abduction is the frontier upside built on a measurable foundation.

Better science also creates a distinctive economic asset. Frontier model development increasingly values expert judgment, executable research decisions, situated trajectories, failures, corrections, and the topology of exploration itself. Individual scientists enter this market with fragmented archives and limited bargaining power. The co-op gives them shared preservation, appraisal, curation, licensing capacity, and collective standing across both existing research artifacts and the forward-looking record generated through the Router.

The near-term proposal builds the co-op and pilots the Router in parallel. Members receive immediate scientific value through better context, criticism, collaborators, preservation, progress, and publishable work. The first economic service offers a free, confidential appraisal of one valuable artifact inside a trusted execution environment; every disclosure and license requires member authorization. The paper invites a founding cohort to test whether this paired institution improves science now and gives scientists durable agency at the intelligence frontier.

Section 1The Research Router and the cooperative proposition

The Router and the co-op solve different parts of the same problem. The Router is the working system: it preserves selected states of inquiry, finds relevant context across a trusted network, introduces complementary investigators, and records what changed. The co-op is the governing institution: it defines membership, quality standards, permissions, representation, appraisal, and the terms under which shared assets create scientific or economic value. The Router gives the co-op a daily scientific purpose; the co-op gives the Router legitimacy, continuity, and collective power.

A researcher using the system

Consider a materials scientist with months of local agent logs from instrument calibration. A backward-looking scan finds a failed temperature schedule whose diagnostic traces could spare another lab a week of work. The local agent drafts a compact contribution describing the setup, failure, and boundary conditions. She edits it, chooses cohort visibility, and publishes it. The remainder of the archive stays local.

During a later experiment, the ambient agent notices that an anomalous sensor reading led her to change the calibration method and immediately suggests a compact state update. She approves it in the flow of work, and it appears on the live Router bulletin board in real time with the anomaly, adjustment, and uncertainty intact. The shared Router can route that live state to researchers pursuing related methods or holding relevant critiques while their input can still affect the experiment. Replication, correction, or reversal becomes a subsequent state transition, preserving both the path of inquiry and its eventual conclusion.

Why this institution matters now

The familiar debate about artificial intelligence and academia asks what machines will automate. Can a model write a literature review, solve a theorem, formulate a hypothesis, analyze a dataset, control an instrument, or draft a paper? Those questions matter, but they are narrower than the institutional change already underway.

A frontier system can serve as research software, fact checker, search engine, development environment, critic, collaborator, and laboratory interface. Deep integration gives it an institutional role resembling the university, journal, laboratory, library, market, and state. It shapes the production of answers and the distribution of attention, authority, capability, and economic power.

This paper uses frontier intelligence to mean the most capable general-purpose AI systems and the increasingly agentic infrastructure around them. The argument begins from a present observation: these systems are becoming capable and integrated enough to mediate a growing share of scientific work. Questions about artificial general intelligence, architectural destiny, and machine autonomy remain outside the claim required here.

OpenAI's academic program provides a sense of the prospective scale. The company plans to provide free frontier-model access to 100,000 scientists, mathematicians, and engineers through 2027, beginning with 10,000 researchers and supplying models, tools, connectors, training, and research support (OpenAI, "Accelerating scientific discovery with ChatGPT for Academic Researchers"). This could be an enormous public benefit. It could also help establish one company's interfaces, workflow conventions, and technical standards as an ordinary environment of research.

That dual character follows ordinary platform dynamics. OpenAI says these workspaces have business-grade privacy and reserves model training on participant and API data for explicit opt-in (OpenAI data controls). The immediate issue is institutional position. Genuine subsidy, superior products, feedback, connectors, and accumulated workflow knowledge can establish a company as the default mediator of scholarly inquiry independent of direct training on customer content.

Academia should welcome tools that expand scientific capacity. It should also recognize the constitutional character of the choice it is making. Who determines the terms on which frontier intelligence reads, learns from, collaborates with, evaluates, transforms, and preserves scholarship? What remains portable when a temporary subsidy ends? Who can inspect the private record through which research decisions were made? And who has standing to negotiate when that record becomes valuable to model development?

The cooperative proposal answers those questions with a causal sequence: the Router makes scientists better at science now; better science creates a uniquely valuable record of inquiry; and the co-op gives scientists collective standing to govern, appraise, license, and bargain over that value.

Section 2From published knowledge to a living record of inquiry

The phrase training data often suggests an undifferentiated pile of text. Contemporary model development uses a richer set of objects, each entering a different stage, performing different epistemic work, and creating different governance problems. These layers form a directional, nonlinear progression:

Published knowledge → expert judgment → executable research decisions → situated trajectories → exploration topology

LayerWhat scientists contributeWhy it matters to frontier models
Published knowledgePapers, books, code, datasets, patents, protocols, and public discussionConcepts, language, procedures, examples, and broad world knowledge
Expert judgmentDemonstrations, comparisons, critiques, rubrics, adjudications, and accounts of what evidence mattersDomain standards for reasoning, behavior, verification, and quality
Executable research decisionsTasks, verifiers, applications, unit tests, environments, simulations, instruments, and experimental testbedsCurricula in which agents can act, receive feedback, and learn what success means
Situated trajectoriesAttempts, interventions, outcomes, failures, corrections, tool traces, and organizational contextEvidence about how performance unfolds in real work and where apparently reasonable procedures fail
Exploration topologyWhich anomaly mattered, who held a minority view, what changed a hypothesis, which path was abandoned, and how a line of inquiry evolvedA record of collective search at the frontier of knowledge, including unrealized branches and decisive negative information

Published knowledge

Academic output is already a deliberate component of important training corpora. The Pile included arXiv, PubMed Central, PubMed abstracts, PhilPapers, FreeLaw, and GitHub among its sources; its authors found that a diverse mixture improved academic and professional subdomains relative to raw web corpora (Gao et al., The Pile; documented components). Newer projects such as the Common Pile and Ai2 OLMo show that substantial models can be trained with more transparent or openly licensed inputs and, in OLMo's case, with much more of the training stack available for inspection.

The exact contribution of academic literature to a current proprietary frontier model remains opaque. Scholarship also reaches models indirectly through textbooks, Wikipedia, code, journalism, citation, and public discussion. The evidence supports a foundational role for scholarship in pretraining while leaving its precise share of any model's intelligence unresolved.

The value of the published stock is now visible in licensing markets. Publishers have begun licensing archives to AI firms. Wiley reported nearly $40 million in fiscal-year 2025 revenue from generative-AI content rights, while Informa reported more than $75 million in first-half 2024 AI-partnership revenue, including access to parts of the Taylor & Francis archive (Wiley FY2025 earnings presentation; Informa 2024 half-year results). These transactions reveal an upstream publisher problem: scholars and public funders produce knowledge, publishers aggregate rights, model companies turn archives into capability, and universities buy access to the resulting systems. Each intermediary may provide value, while researchers and learned societies often sit outside the negotiation.

Expert judgment and executable research decisions

Pretraining gives a model broad patterns. Post-training teaches it how to behave, which answers are preferred, how to use tools, and what counts as success in a domain. In verifiable areas, models can generate synthetic practice and learn from objective checks. Unit tests give programs binary feedback, and proof assistants can mechanically verify some formal proofs. Where quality is ambiguous, situated, or expensive to observe, human judgment remains harder to replace.

This is why a referee report is a different kind of model input from a paper. The paper presents a conclusion. The report exposes why an argument is inadequate, which alternative explanation has been ignored, which evidence would discriminate among mechanisms, and what standard the work must meet. A seminar records challenge and repair in real time. A failed experiment reveals where an apparently reasonable procedure breaks. A good benchmark or verifier turns some of this judgment into an executable test.

Mercor's public description of its business is unusually direct. It mobilizes subject-matter experts to create specialized data, benchmarks, and evaluation environments. Its reinforcement-learning pipeline builds data-rich worlds, implements the applications and tools agents need, and constructs rigorous tasks and verifiers (Mercor Research). In a company-published experiment, 874 expert-authored tasks across 50 simulated enterprise worlds improved an open-weight model on held-out professional tasks. As vendor-produced evidence, it supports the economic thesis that hundreds of carefully designed examples may be more useful than an enormous unstructured archive while leaving independent validation to future work (Mercor, "Expert data drives model performance").

Mercor CEO Brendan Foody describes the bottleneck as organizing human intelligence. In his account, useful post-training material consists of tasks containing prompts, verifiers such as rubrics or unit tests, and sometimes environments; converting tacit knowledge into verifiable rewards is itself the data (Foody, June 11, 2026; Foody, Aug. 3, 2026). Foody also argues that synthetic-data businesses struggle when models have limited ability to identify the frontier of their own mistakes (Foody, Aug. 8, 2026). Humans can also be inconsistent, orthodox, or wrong. The durable point is that the scarce input in many domains is a consequential choice about which problem to attempt, what a good solution would look like, and which evidence should change our mind.

Tools, environments, and synthetic laboratories

The word tool should be read broadly. It includes code interpreters, shells, virtual machines, databases, search systems, continuous-integration systems, automated differentiation and optimization loops, sensors, instruments, simulations, and physical or computational testbeds. An environment may be as modest as a repository plus a shell or as rich as a simulator linked to a real instrument.

This matters because agents can already interact experimentally with the world. A coding agent that writes a program, inspects compiler output, queries a database, runs a simulation, or changes a parameter in response to sensor data takes an action, observes a result, and revises. The synthetic laboratory already exists as a widespread workflow in software and systems research, while richer sensor and instrument loops extend the same pattern into physical science.

The strategic object is expanding from text, to judgment, to executable curricula, and then to situated research trajectories. Evidence connecting arbitrary agent sessions to safe, durable general capability remains preliminary, and continual-learning and self-distillation results remain mixed (SDFT; u-OPSD; Denser study). The option value matters. A platform that mediates scientific work today may control uniquely valuable streams of situated experience tomorrow.

Section 3A testable scientific case: recall, precision, and the occasional jump

The co-op earns its place by improving ordinary research activity, with occasional breakthroughs as additional upside. Its near-term scientific claim is that routing can increase recall and precision. These terms describe the quality of timely scientific context, distinct from popularity or output volume.

Recall and precision support all three familiar forms of reasoning. Induction improves when researchers see a broader and less redundant set of observations, including failures that often disappear. Deduction improves when premises, definitions, and chains of evidence can be inspected and contested. Abduction becomes more likely when distinct investigations preserve different premises long enough for a productive bridge to form. These processes interact: a new frame generates consequences; testing those consequences then strengthens, revises, or rejects the frame according to evidence.

Tom Zahavy's position paper on abductive reformulation argues that contemporary language models are powerful at induction and increasingly capable at deduction while genuine abductive reformulation remains difficult. He proposes physically grounded world models as one route toward manipulative abduction (Zahavy position paper). The position identifies an important difficulty. Models can generate surprising hypotheses, and coding-agent workflows already support experimentation through execution, observation, and revision. The cooperative thesis rests on the additional search breadth and contextual variety that networks provide.

Testable social hypothesis. A partially connected network of strong, distinct human-agent investigations that routes problem-relevant research states among them can improve awareness of relevant work, reduce avoidable duplication, strengthen criticism and verification, create useful collaborations, and occasionally enable new explanatory frames relative to isolated investigations.

The claim is deliberately about a partially connected network. Selective connectivity preserves investigative autonomy while targeted bridges carry relevant context. This design contains coordination costs, limits fashion-driven convergence and status effects, and lets researchers treat every routed encounter as an opportunity they may evaluate on its merits.

Several bodies of evidence make this proposition worth testing. High-impact work often combines a conventional foundation with atypical knowledge (Uzzi et al., "Atypical combinations and scientific impact"). Smaller teams are associated with more disruptive work while larger teams tend to develop existing trajectories, suggesting that a healthy ecology needs both (Wu, Wang, and Evans, 2019). In cumulative cultural innovation experiments, partially connected groups preserved more variation for later recombination than fully connected groups (Derex and Boyd, 2016). Recent bibliometric work likewise suggests that bridges across structurally distinct prior communities can support disruptive work, although such metrics are imperfect and best used to motivate experiments while the question remains open (Peng et al., 2026).

The practical consequence is a broader account of scientific quality. A useful Router serves generative and critical researchers alike. Some members produce sparks, analogies, and alternative mechanisms. Others are exceptional filters: they identify unmet standards, missing controls, prior failures that make a path unpromising, and polished output that lacks substance. A persistent critic that remains quiet until a known objection becomes relevant may be as valuable as a collaborator who proposes a new direction.

Section 4What the Router routes

The Router is a scientific routing layer built around selected research state or state transition: a question, hypothesis, result, failed approach, blockage, method, objection, request for help, piece of evidence that changed a view, or proposed next test. Members choose these compact, purposeful records in place of raw chain-of-thought. A useful entry carries more scientific context than a social post and circulates faster than a paper, preserving enough context for another person or agent to contest, avoid, reuse, or extend the work.

Routing combines four functions: taking notes on active inquiry; searching a shared record and injecting relevant context; communicating outward through notifications, digests, or drafted responses; and facilitating introductions or proposed collaborations. In some cases the output may be a person to call. In others it may be a null result, a method, a criticism, or a draft synthesis between two live investigations.

Ambient publishing is the Router's live suggestion loop. The local agent notices meaningful changes in an authorized workspace and drafts compact state transitions that may be worth publishing. The researcher edits, approves, and chooses the audience for each contribution; unused suggestions remain local. Approved states become searchable immediately, so the network can return context while a question is active.

A first-use experience

  1. Inspect locally. A local Router agent scans the researcher's own journals and agent logs on the local machine, maps the private iceberg beneath finished outputs, and explains what may be valuable to the member or community.
  2. Review live publication suggestions. As work progresses, the local agent drafts meaningful state transitions that the researcher can edit, publish to a chosen audience, or leave private.
  3. Search during active work. When a live question, method, or blockage appears, the agent searches the co-op's record for related states, including structurally relevant work expressed in different language.
  4. Return grounded context. The agent explains why a result, objection, or person may matter, includes provenance, and distinguishes source material from its own synthesis.
  5. Ask before disclosure or contact. The system can draft a contribution, question, or introduction. The default sends each outbound action to the owner for approval; members may also establish narrow pre-authorized rules.
  6. Preserve the transition. If the encounter changes the work, the local agent records what was learned, rejected, or attempted next and suggests a publication for the researcher's review. An approved transition becomes searchable for future investigations.

A three-layer architecture

LayerPrimary locationFunctionDefault disclosure rule
Shared RouterCo-op networkPublish and search selected research states; deliver context, digests, critics, and introductionsOnly contributed or otherwise authorized material is visible
Local Router agentMember-controlled device or environmentInspect private archives for useful prior work; notice live state transitions; suggest posts and responses during active workThe researcher approves content and audience before publication
Trusted executionTEE or comparably bounded computeMatch, search, test, or appraise protected artifacts while the corpus remains protected and in placeReturn bounded answers, evidence, or offers under owner-approved rules

The architecture makes the bulletin board a live surface and a pointer to something deeper. Researcher-approved ambient publications keep it current while the complete private record remains local. A member may contribute a result and retain the surrounding scripts, traces, failed branches, and confidential material. Later, a trusted query can ask whether those private materials contain something relevant and return an owner-authorized answer.

The persistent-critic pattern is especially important. Some knowledge should be queried many times and returned rarely. A member may authorize an adviser to compare a live plan with a protected library of objections, safety conditions, null results, or disciplinary standards. Most checks remain quiet. When a known failure mode becomes relevant, the agent surfaces the critic and explains the match. This imports critical context while private notes retain their intended scope.

A pre-preprint layer

The Router should host scholarly objects between a private notebook and a finished preprint. A useful object might preserve the question as posed, the intervention attempted, the result obtained, the assumptions challenged, and the next discriminating test. It should be available while the question is still alive, with enough structure for correction, withdrawal, citation, and continuous peer review of the transition.

A small HypoArena experiment conducted during preparation of the earlier draft illustrates the form of a routed state transition. The intervention was underpowered and produced a null result. The Router preserved the conditions of failure and a proposed next test in a form another researcher could contest, reuse, or redesign while the question remained open (HypoArena/HypoData; Router entry). The example shows how a failed attempt can become living scholarship; the broader network thesis remains for the pilot to test.

Section 5Governed discoverability

Researchers often keep unfinished work private because sharing carries a bundle of concrete risks:

Governed discoverability addresses these risks through private, cohort, and public layers; local-first inspection; selective contribution; privacy-preserving answers; owner notification and drafted responses; provenance; correction and withdrawal; and permissions tied to purpose, recipient, action, and time.

These permissions make ambient publishing practical. A member can choose which workspaces the local agent observes, what kinds of changes it should suggest, and which audiences are available. Each suggestion remains local until the member approves its content and audience. The local Router agent attaches provenance and policy metadata to every published state, and members can correct, withdraw, or narrow access as the work develops.

TinyCloud's strategically important primitive is trusted execution: a bounded program can operate against protected material and return an authorized result while the underlying corpus stays protected and in place. The operating principle is protected query. A co-op could ask whether two private archives contain complementary methods, whether a candidate artifact appears relevant to a buyer's evaluation, or whether a protected critic matches a live plan, while revealing only the bounded answer and evidence the owner authorized. TinyCloud's capability model can also narrow permissions by action, path, expiry, and delegation (TinyCloud capabilities).

Trusted execution is a timely asymmetric advantage for a cooperative because many incumbent data workflows copy material into a buyer or intermediary's environment before evaluation. TEEs supply a technical boundary; governance supplies legitimacy and remedy. Membership rules, contracts, security review, audit, legal diligence, sanctions, representation, and credible remedies determine permitted uses, voting power, value allocation, code assurance, side-channel response, and treatment of information embedded in trained weights.

The co-op should also distinguish public scholarship from private and pre-publication research objects. Published claims and public evidence should remain broadly available for scrutiny, education, replication, and unexpected reuse. Open, machine-actionable, authorized, and reciprocal describe different properties. UNESCO's open-science framework emphasizes community-owned infrastructure, interoperability, preservation, shared benefit, and safeguards against inequitable extraction (UNESCO Recommendation on Open Science). The FAIR principles anticipated computational agents as stakeholders by asking that scholarly objects be findable, accessible, interoperable, reusable, licensed, and provenance-rich enough for machines to know how they may act (FAIR Guiding Principles).

Reuse requires explicit terms alongside open access. Most arXiv papers are freely readable, while license terms vary and bulk access has separate operational rules (arXiv license information; arXiv bulk-data guidance). The CARE principles, developed for Indigenous data governance, further place collective benefit, authority to control, responsibility, and ethics alongside technical openness (Global Indigenous Data Alliance, CARE Principles). A cooperative can maximize scientific distribution where openness is appropriate while governing private, sensitive, and pre-publication objects on terms that sustain future participation.

Section 6Two asset classes, one causal sequence

The economic object has a backward-looking and a forward-looking form. The distinction is useful because it allows the cooperative to begin before the full Router network exists while keeping its north star clear.

Backward-looking: the record scientists already possess

Scientists, laboratories, and research companies already hold private journals, agent logs, code histories, failed experiments, instrument traces, annotated papers, internal critiques, workflows, and experimental archives. Much of this material is slightly illegible because it was produced to do research. That is part of its value. It records how exceptional practitioners investigate, choose, fail, revise, and decide what evidence matters in addition to the conclusion they later publish.

The first obligation is preservation. Traces that expire take future understanding, appraisal, routing, and bargaining options with them. A local Router agent should therefore make retention settings visible, warn when potentially valuable records are expiring, support local export in durable formats, and help members compress or selectively retain material according to its option value. The call to action is modest but urgent: preserve the option while the market, governance, and scientific uses are being understood.

Local inspection also creates immediate feedback. A scientist should be able to see how her archive is distributed across topics, methods, tools, collaborators, failures, and unresolved questions; which fragments might be useful to her community; and which public entries could serve as the tip of a deeper private iceberg. Appraisal begins as self-understanding before it becomes a transaction.

Forward-looking: the record the Router creates

Researcher-approved ambient publications let the Router create a higher-entropy record of live inquiry: state transitions selected because they helped someone make progress, avoided duplication, imported a critic, changed a hypothesis, or produced a new collaboration. Its value can compound. Each well-routed contribution improves the network's future recall and precision; each correction makes later synthesis more trustworthy; and each preserved branch lets other agents explore novel parts of the landscape.

The forward-looking record gains value from useful information per unit of attention and reasoning. Repetition, slop, and contextless output lower that value. Provenance, disagreement, null results, and evidence that changed a view raise it. The cooperative should optimize for the quality and topology of the record, using gross capture only where it serves that aim.

Over time the boundary of the record may move closer to cognition. Agent-mediated work already produces a higher-resolution account than papers or version-control commits. Future interfaces may capture richer interactions, and brain-computer interfaces may eventually create still more situated traces. This horizon strengthens the case for designing member agency into the institution now, so increasing resolution serves informed participation and shared benefit.

The two asset classes reinforce one another. Existing artifacts let the co-op provide preservation, appraisal, and bargaining services immediately. The Router gives members scientific reasons to participate and creates a more valuable future record. Early services can begin with a small corpus, while every economic service remains anchored to the scientific network that makes the material worth having.

Section 7Build the cooperative and pilot the Router in parallel

The protagonists of this paper are scientists choosing to build a selective cooperative that behaves partly as a learned society, partly as a library, partly as a standards body, partly as a data trust, and partly as a live research network. These functions can begin across several aligned entities joined by a shared compact and interoperable infrastructure.

Selectivity is central. The quality of the people and investigations is part of both the scientific product and the bargaining asset. A first cohort should recruit leading practitioners of agentic science across disciplines who are doing strong, distinct work, who can articulate their standards of evidence, and who are willing to contribute useful states while retaining local autonomy. A learned-society quality bar should sustain trust and produce genuinely valuable differences, with contribution quality and complementarity carrying more weight than prestige.

What a member should experience first

This sequence matters. The Router first helps scientists understand their own archive and make better progress. Members then choose contributions from a position of informed trust and visible benefit. A bulletin-board post becomes an invitation to a deeper scholarly exchange, with the surrounding scientific record available through governed pathways.

The co-op and Router should therefore be built in parallel. The co-op supplies membership, trust, rights diligence, representation, and bargaining authority. The Router supplies ambient scientific exchange and a forward-looking record that makes the institution more valuable. Existing artifacts make the co-op useful before the network reaches scale; live routing keeps the institution oriented toward future scientific value.

Why a cooperative

Model developers and data intermediaries create value through engineering, compute, screening, standardization, quality control, product design, and risk. Scientists create foundational knowledge, judgment, and research trajectories. A healthy market gives each group access, information, portability, voice, and a share of the value it helps create.

Individual bargaining is poorly matched to model development. A single trace may have little value alone and enormous value in combination. The buyer usually knows more than the worker about downstream reuse, substitutability, destination, and licensing life. Attribution becomes harder after assets are mixed into datasets, environments, and training runs. A cooperative can aggregate supply, curate quality, create demand, conduct confidential price discovery, and negotiate a framework across transactions on behalf of scientists acting together.

Members should aspire to own and govern both the institution and the bundle of collective rights it represents. The founding process will determine the legal form, voting rules, representation, ownership transition, scientific admissions process, rights diligence, ethical exclusions, security standards, audit, correction, deletion, sanctions, and revenue distribution. Naming these questions now gives the founding cohort a concrete constitutional agenda.

Relevant operating precedents

The proposal joins institutional capacities that already operate in narrower settings. The examples below show that member-governed data access, cooperative research infrastructure, community ownership of scientific intellectual property, and sector-wide permission systems can each function in practice.

PrecedentDemonstrated capabilityWhat this proposal adds
[MIDATA Cooperative](https://www.midata.coop/)Member-governed, consent-based access to personal health data for researchScientists' working traces, live routing, and collective model-market negotiation
[Red Hen Lab](https://www.redhenlab.org/welcome-to-potential-contributors)Researchers pool archives, tools, compute, and labor in a distributed laboratoryFormal member governance, ambient publishing, protected appraisal, and economic services
[VitaDAO](https://www.vitadao.com/research)Community funding, ownership, and commercialization of research intellectual propertyDaily scientific routing, governed private archives, and the process record of inquiry
[JoinData](https://join-data.nl/en/)Sector-wide infrastructure for owner-authorized access to farmers' business dataPermissioned research states, agent logs, and collective evaluation or licensing

These precedents demonstrate that member-governed data access, cooperative research infrastructure, community ownership of scientific intellectual property, and sector-wide permission systems can each operate in practice. The Research Router Co-Op joins these capabilities around live, agent-mediated inquiry: ambient publication, governed access to deeper research archives, scientific routing, confidential appraisal, and collective negotiation over the resulting record.

Emerging example. Africa's Voice describes contributor ownership, consent tiers, and dataset-sale dividends for AI-training data. Its current public materials present a prelaunch model: the contributor page labels the app "available soon," and the team page currently contains placeholder names (accessed Aug. 25, 2026).

Section 8Appraisal, licensing, and collective bargaining

The first economic service should be narrow enough to trust and useful enough to learn from: a free, confidential appraisal of one valuable scientific artifact inside a trusted execution environment. The member might submit a research journal, agent-log segment, benchmark, verifier, dataset, method archive, or record of failed experiments. Member authorization governs every transfer, publication, and license. The service returns a bounded assessment of possible scientific uses, potential buyers or collaborators, rights and sensitivity concerns, comparables, and evidence of demand suitable for safe disclosure.

Appraisal is a governed evidence process and an open mechanism-design problem. Value depends on quality, rarity, verifiability, rights, complementarity with other assets, the model and objective, the cost of producing substitutes, and the scale of downstream reuse. A TEE can reduce the information asymmetry of negotiation by allowing a buyer's evaluation logic and a member's protected artifact to meet under agreed rules. Human judgment, competing bids, disclosed assumptions, and conflict rules remain essential because valuation stays contingent and interested.

The cooperative's first market proposition should focus on private research journals and agent logs because they contain granular evidence of how exceptional scientists investigate. Demonstrations, critiques, rubrics, tasks, verifiers, environments, experimental archives, null results, and situated trajectories form additional asset classes that may be easier to evaluate or license in particular domains. The co-op should develop common appraisal frameworks by asset type, moving beyond a single price per token or hour.

Collective bargaining includes pooling supply, curating a coherent corpus, establishing quality and rights standards, shopping bounded descriptions to multiple counterparties, comparing demand, creating new products, and coordinating where members explore. Aggregation changes both the amount of information available and the conditions under which it can be evaluated. When the scientific network improves the underlying work, the co-op helps create more value before negotiating a fair division of it.

Plausible mechanisms with calibrated precision

The co-op should describe attribution with calibrated confidence. Fine-grained provenance can record who introduced a claim, supplied a method, changed a hypothesis, or ran a decisive test. Each person's marginal contribution to a future model capability remains a harder scientific and economic question. Process attribution can support credible payments and rights today while research on capability attribution continues.

DAOs, decentralized science, and public-goods funding offer adjacent mechanism families that can complement confidential appraisal, collective negotiation, and member governance. Buterin, Hitzig, and Weyl's Liberal Radicalism illustrates one family of plural funding mechanisms. MaxShapley offers evidence that bounded document-level attribution and retrieval valuation can improve. Both advance useful mechanisms while leaving the exact division of value for emergent model capability open.

Recent economic work gives formal support to the bargaining intuition. Labor as Capital argues that individual ownership of work-generated expertise data can cause substitutable workers to compete away their share of AI productivity gains, while collective ownership can improve both knowledge supply and the division of value (Cullen, Li, and Li, 2026). Historical experience with producer-governed dairy data shows that collective stewardship can support large-scale research while preserving producer control over a critical asset (Hutchins et al., 2023). These analogies motivate tests of institutional forms while leaving the legal design to empirical and constitutional development.

Section 9A pilot that can fail

A credible first pilot should test the scientific and economic propositions separately enough to learn from failure while allowing them to reinforce one another. An ARIA-relevant cohort is a useful setting because its researchers already pursue distinct, high-agency work and have reasons to preserve the methods, tools, and failed branches beneath finished outputs.

Six stages

  1. Preserve. Help participants inspect retention settings, export journals and agent logs, identify sensitive material, and keep valuable records locally.
  2. Index. Build a local map of questions, methods, tools, experiments, failures, objections, and candidate contributions while the full archive remains on the member's system.
  3. Suggest, publish, and route. During active work, offer participants ambient publication suggestions; publish each approved contribution to the shared Router immediately; then surface relevant co-op entries and protected matches with explanations, provenance, and confidence sufficient for the researcher to judge usefulness.
  4. Connect. Offer owner-approved introductions, requests for critique, or collaboration proposals when two live investigations appear complementary.
  5. Measure. Compare routed and baseline periods or tasks using operational measures tied to scientific usefulness, attention cost, and changed decisions.
  6. Appraise. Invite each participant to submit one artifact for a confidential TEE-based appraisal under member-controlled licensing, then learn which evidence of value and demand is actually useful.

Scientific measures

Recall has an unobservable denominator: all relevant work that could have helped. The pilot can estimate it through tractable proxies and blinded review:

The pilot should record the costs as carefully as the benefits: interruptions, confidentiality incidents, social pressure, prestige bias, duplicated notifications, false matches, performative posting, and premature convergence. A Router that increases nominal recall by flooding attention has failed. A selective network that produces fewer but more consequential encounters may be healthier than a busy one.

Economic and institutional measures

Baselines, sample size, success thresholds, admissions criteria, rights diligence, and the boundary between automatic privacy-preserving answers and owner approval should be set with the founding cohort. The pilot should support genuine falsification. Results may show that routing is too noisy, appraisal too speculative, trusted execution too cumbersome, or the cooperative proposition insufficiently compelling. Those findings would redirect the institution before its defaults harden.

Section 10The wager

The wager behind a scientific cooperative is that intelligence – human or artificial – develops differently depending on the network in which it participates. An isolated agent may be extraordinarily knowledgeable and still search inside the wrong frame. A collection of identical agents may amplify the same blind spot. A fully connected network may converge too rapidly on the first persuasive explanation. A partially connected network of strong, distinct investigations can preserve different premises, methods, experiences, and standards long enough for evidence or a productive bridge to matter.

The aim is to keep human scientific communities active in choosing questions, judging evidence, preserving disagreement, and sharing in the capabilities that result. Frontier agents can increase the amount of inquiry that is attempted. The Router can help distribute that search across a broader landscape with less redundant effort and better critical context. The co-op can give the people generating the record standing to govern its use.

If the wager is right, the institution produces three kinds of value at once. It improves the recall and precision of scientific work now. It gives scientists collective leverage in the emerging market for judgment and situated research data. And it creates a governed scientific memory from which future agents can learn both how inquiry changes and what polished conclusions look like.

Academic knowledge and research workflows will continue to circulate. The institutional choice concerns who establishes the terms: publisher licenses, contractor dashboards, proprietary research interfaces, and platform policies, or member-governed structures that give scientists voice, access, obligations, portability, and collective power.

The practical invitation is therefore twofold. An individual scientist should be able to apply to the co-op with one valuable artifact, receive local preservation and appraisal support, and choose each contribution while the rest of the archive remains protected. A cohort or research leader should be able to pilot the Router, measure whether it improves recall and precision, and introduce the cooperative proposition to researchers through evidence and lived experience.

The Router should make scientists better at science now. Better science should create a valuable record of inquiry. The co-op should give scientists the standing to govern, appraise, license, and bargain over that value. A founding cohort can write the constitution while the technical and economic defaults remain open.

Evidence and status note

Sources were checked through Aug. 13, 2026; linked material carries later dates where indicated. Company publications and executives' posts document how firms describe their strategies; independent validation of performance and scale remains a separate evidentiary category. Bibliometric studies are qualified where citation impact or disruption serves as a proxy for scientific value. Descriptions of the Router, local Router agent, ambient publishing, TinyCloud, appraisal service, selective co-op, pilot, and BCI horizon are project descriptions, design proposals, and hypotheses awaiting empirical tests. The paper invites a founding process; participant materials remain governed by owner authorization and future co-op rules.

If the argument holds for you, put your name to it.

Add your name →