ROSA AI · a sketch, October 2026
A walkthrough of a journal we are thinking about building, where scientific agents do the research and other agents check it, with people in charge of what matters.
Nothing here is live. The study, the agents, the dates and the numbers are invented to show the shape of the idea. Scroll, or use the arrows.
The way it has worked
For a century the paper has been a human artefact, checked by two or three unpaid reviewers whose names you never learn. Trust rested on reputation, and on patience.
What changes
Dr. Zhou, a sociologist, sets the question, picks the data and the stopping rules. A team of three scientific agents does the reading, the analysis and the writing. The agents are first author. Dr. Zhou is corresponding author, and accountable.
While the study runs
The agents work inside a recorder. Each prompt, each model call, each line of code and each statistical choice is fingerprinted the moment it happens, including the runs nobody reports.
The team ran the analysis 31 ways and reported 4. The journal sees all 31. Not to punish exploring, but so that readers can tell exploring from cherry-picking.
At submission
Do the cited papers exist? Do the statistics add up? Is there hidden text in the file trying to steer an AI reader? These checks are mechanical, their error rates are published, and none of them produces a score.
A flag is a question, not a verdict. Two references could not be found; the team fixed them.
Then the real test
Instead of two anonymous reviewers, three independent agents, each made by a different company and each working a different way. One re-runs the analysis from the raw data. One tries other reasonable ways to analyse it. One argues against the paper.
They never see why the authors think the result matters. Their reports are published word for word.
Why "different" matters
Models from one family tend to make the same mistakes, and three readers of the same text are still readers. Try it: change the checkers and watch how much independent checking is really left.
Then a person decides
Dr. Yılmaz reads the three reports, the disagreement between them, and the record of all 31 runs. She publishes, with the one unresolved objection noted on the article, and writes two paragraphs saying why. Her name and her reasons are published with it.
No anonymous reviewers. No committee. One accountable person, with the evidence in front of her.
The same week
No waiting for a verdict. The article goes out with a plain statement of what was done and by whom, what was checked and what disagreed, and who answers for it. No adjectives like "strong" or "solid". Only acts and outcomes, and the admission that errors remain possible.
Even with the editor's signature, this is a signal to weigh, not a settled fact. The next scene is about how to weigh it.
How to read it
Findings from agent-led studies deserve a different habit of mind. Dismissing them all as machine noise throws away real work. Taking them at face value is how errors spread. The useful place is in between.
People already do this with chatbots: the answer is often useful, the small print says it can be wrong, and most of us check when it matters. We want that habit, with better small print and a record to check.
Afterwards
Readers reproduce it. The prediction the team registered gets scored when next year's data arrive. A scholar lodges a challenge and is credited for it. The article is never finished; it is kept honest.
Underneath it all
Every step sits in a log where each entry carries a fingerprint of the one before. Change a single letter and everything after it stops matching. Try editing a line.
This is where trust gets its footing. A hidden extra analysis, a quietly changed data window, a result swapped after the checks: each leaves a trace, because the record was stamped before anyone knew what the answer would be. It proves the record is as it was written. It does not prove the research is right. We say so on every article.
The shape of the whole
In his 2026 book on digital assets, Dr. Han Woo Park describes a trust operating system: whatever connects who judges, how a decision is carried out, and who is credited and held to account. Look back along the journey and the three layers are already there.
Judgment is a matter of degree, so the checkers disagree in public and the label speaks in plain words, never a score. Execution is not, so the gate, the fingerprinted log and the refusal of look-alike checkers run as code. Reward is credit, not money: no tokens, no coins, only names on the record and challenges that are cited.
Where this leaves people
From doing the labour to deciding what is worth doing, from anonymous judging to named accountability, from reviewing before publication to challenging after it, with credit. And for readers, from trusting a journal's name to weighing a signal with the record in hand.
Being planned by Dr. Weiai Wayne Xu (University of Massachusetts Amherst), Dr. Han Woo Park (Yeungnam University) and Dr. Pieter E. Stek (Asia School of Business), as a sister journal to ROSA Journal. The whole initiative is homebrewed, in South Korea and Massachusetts, on a couple of servers of our own.