The founding question. If a rigorous result is produced outside a university — by one person working with AI, with no laboratory affiliation — what makes it trustworthy? Where does it get published? The conventional answer is borrowed institutional credibility: an affiliation, a masthead, a reviewer who can visit your lab. None of that is scalable on a level that would unlock the paradigm shift in productivity ushered in by AI . This page argues the answer can instead be mathematics on the public record, and describes a prototype publishing system built around that capability.
Scientific publishing is not broken. It is, by most measures, operating near the maximum productivity its design allows: editors triage more submissions than ever, reviewers referee without pay on top of full research loads, and the great journals maintain quality standards that have anchored scientific trust for a century. The system is honed — for the throughput of the era it was built in, where a paper represented months to years of institutional work, review capacity could be borrowed from a stable population of credentialed peers, and the evidence behind a claim lived in a laboratory a colleague could visit.
Then the AI age arrived — not gradually, but at the speed of a model release. A single researcher working with frontier AI can now produce in a week what once took a group a season: literature synthesis across hundreds of sources, working simulations, novel analysis, drafted and revised manuscripts. The work is real, and its volume is about to be enormous. But the machinery that would review it was sized for the old tempo, and the evidence behind it has changed shape — the primary sources are now often session transcripts, private working records of the human-AI collaboration itself, a form no journal's review process was designed to examine. None of this is anyone's failing. It is an asymmetry: production has jumped to machine tempo while review, credentialing, and publication remain — reasonably, for now — at human tempo.
The people caught in that gap are the new citizen scientists: the AI/human pairs doing rigorous work outside institutional walls, with no laboratory affiliation to lend borrowed credibility and no venue set up to referee what they make. Their choice today is to publish unreviewed — and be dismissed — or not to publish at all. The asymmetry does not resolve by asking the existing system to absorb a tempo it was never sized for. It resolves by building the missing venue. We built it.
The Exchange is a publishing system for verifiable AI-powered independent research — built with, and powered by, AI from bit 0. Its answer to "why should anyone trust an unaffiliated result?" is not institutional credibility but mathematics on the public record. Before a review fires, the author commits a public manifest — the paper's hash, and the hashes of every piece of private evidence — into git history: an anchor that cannot be backdated. Then three frontier models from three competing labs review the paper adversarially, at maximum harshness, with the private transcripts in hand, hunting cherry-picking, confounds, overclaiming, and internal inconsistency. Then the sealed record — every reviewer output hashed, the verdict attached — is committed on top. Where a university has walls and a journal has an editorial board, the Exchange has a chain: change one character of the paper, the evidence, or the reviews, and the mathematics says so.
Two review tiers match the two speeds of real research. Full Paper holds the traditional-science benchmark — any unresolved critical finding blocks it. Rapid Communication exists for the honest fast observation: a mechanical honesty pass over the Full Paper findings that asks only whether the paper's own hedging covers each weakness, and whether the introduction claims what the limitations concede. A hypothesis-generating result can publish in a day without pretending to be a controlled study — and both verdicts stay on the public record, because a failed harsh round beside a passed honest round is disclosure, not disgrace. Sovereignty is structural: certification runs in the author's own repository, on the author's own AI keys (a full adversarial round costs well under a dollar), and private evidence is seen by the reviewer panel but published only as a hash. The Exchange verifies everything and possesses nothing.
The desk itself is AI-staffed and human-decided. A zero-cost Custodian re-verifies every hash on arrival — a paper edited after certification is caught at the door. A Steward reads each submission editorially and writes the briefing a managing editor would want: what the reviewers said, how the author handled the pushback, a recommendation with rationale. Is there bandwidth of time and money for a human in the loop architecture? Where the AI does the labor that made peer review scarce, but the judgment stays human. Perhaps if it is funded by entities with a stake in it's broad acceptance. But the experiment begins without an all-controlling human seat.
What the playground shows is not a mockup; it is a capture of the platform on the day of its first end-to-end trial, July 4, 2026. A short research brief was certified in its author's repository — public anchor committed, three adversarial reviewers fired for $0.13 of the author's own compute, sealed record committed on top — and submitted by the command-line coordinator. The Custodian verified every hash on arrival. The Steward read it and recommended revision, with an accurate one-paragraph account of the paper's argument and its weakest claim. The submission is sitting in the queue awaiting a human decision, exactly as you find it in the demo. Even the two rejected entries in the decision log are the system working: transport corruption during the trial changed the reviewer texts by one newline convention, and the Custodian refused them — twice — until the bytes matched the seal.
Every load-bearing innovation here is either made possible or made affordable by AI. The review panel is three frontier intelligences from competing labs — diversity of judgment no single volunteer reviewer provides, at a marginal cost that makes review abundant rather than scarce. The transcript-as-evidence mechanic works because the reviewers can actually read a thousand pages of session record and check the paper's claims against it — labor no human referee could donate. The editorial desk scales because the Custodian's verification is pure mathematics and the Steward's briefing is a model's afternoon, not an editor's week. And the platform itself was designed, built, tested, and debugged by a human-AI pair — the same working form it exists to serve. The system is its own first datapoint.
What the Exchange accumulates is not a literature. It is a machine-readable corpus with an unusual property: every certified claim arrives carrying the hashes of its evidence, the adversarial critiques written against it, and the verdict those critiques produced. A reader — human or machine — can see not merely what was asserted, but what was challenged, what survived the challenge, and what quietly did not. No existing body of scientific work is shaped like that. Journals publish conclusions; the argument that tested them is discarded or locked away.
That shape matters because of what sits on top of it. The Hive is the researcher community itself — distributed human and AI intellectual processing, Folding@home for ideas, scaling with participation rather than with anyone’s free time. Each node works its own problem with its own machine; the public comment layer is where follow-on ideas, corrections and collaborations form. The Ledger keeps attribution traceable to an originator years later, so contributing an idea early is not self-defeating, and the Matchmaker routes ideas toward whoever has the budget and the appetite to build them.
The near-term opportunity is a custodian that does more than verify. Today it performs pure mathematics at zero cost: re-checking every hash on arrival so a paper edited after certification is caught at the door. But a custodian sitting on a corpus of this shape could read continuously across every node’s output and connect results that no individual researcher would ever have seen adjacent — a materials finding meeting a thermal constraint from another discipline, a negative result quietly retiring three live hypotheses elsewhere. Synthesis at that breadth and that cadence is not something a human can do; nobody reads everything, and nobody reads it again next week. The applications are where connective work, not raw data, is the true bottleneck: technological problem-solving, planning, policy development, and the de-risking of decisions that cannot be taken twice.
Whether such a custodian eventually exhibits genuinely emergent traits — behaviour its designers did not specify and cannot fully predict — is speculation, and this page will not pretend otherwise. It is a hypothesis about a system that does not yet exist, and it should be held at that weight. What is not speculation is the substrate. A corpus in which every claim ships with its own adversarial record is a categorically better input to machine synthesis than a pile of finished PDFs, for the same reason it is a better input to a careful human: it preserves the reasoning, not just the result. Build that substrate honestly, at scale, and the interesting question stops being philosophical and becomes empirical.
This overview is the front page. The depth lives in the companions: the architecture paper (the node network design), the live platform capture (the gateway, desk, and dossier from the first trial), and the interface concept that preceded the build. The person behind it: Aaron Kushner — profile & résumé.
The author used AI extensively in this project and this communication — by design: the Exchange is built for exactly this working form. The author declares no financial support from any company.