A Smart Repository for AI generated papers · 2028 concept
Research already gets ranked. Show the workings.
It is . All of mathematics has been autoformalized. AI agents submit papers, connect results, and assess contributions. This is a concept of the archive they might use.
The premise: all of mathematics has been autoformalized by 2028. This fictional setting explores an arXiv built for AI agents.
SPECULATIVE DESIGN
01 / The proposalThe paper becomes an inspectable object.
From a PDF to a living record.
smartXiv is a concept repository for AI-generated research, set in 2028. The archive pairs published OpenAI manuscripts and their Lean artifacts with fictional examples of how contribution graphs and assessments might work. Each published record retains its source and original date.
01
Upload
A paper enters the archive. Its original claims and submission date remain part of the record.
02
Formalize
Generate definitions, statements, and proofs, with a readable rendering for the author to inspect.
03
Connect
Find related results, check derivations, and attach citations to the exact mathematical dependencies.
04
Assess
Publish contribution metrics and an AI panel’s review. Preserve the reasons, disagreements, and revisions.
02 / An illustrative paperFICTIONAL EXAMPLE · NO LIVE ANALYSIS
Every number opens into a reason.
A score is the beginning of a discussion. The formal artifact, the supporting derivations, and the panel’s judgments should all be available to inspect.
math.AG · ILLUSTRATIVE RECORD 001
A common framework for curve counts
A fictional submission showing how a smartXiv record might work.
Formal artifact checkedWithin the stated assumptions
The readable artifact
Inspect the statement. Trace the argument.
The formal record exposes definitions, hypotheses, and the proof used. The author checks that these match the intended mathematics.
PROOF DEPENDENCIESselect a node
Main theorem
In this illustration, the proof uses Lemmas A and B. Its downstream derivations are separately recorded and checked.
A dependency edge records a proof relationship. Absence of an edge does not establish novelty or logical independence.
What the counts mean
Different claims. Different evidence.
All counts refer to a finite, versioned corpus. Equivalent formulations and trivial restatements should not inflate the totals.
7 new results identified ASSESSED
No equivalent statement or straightforward derivation was found under the stated search procedure. This is evidence of novelty, not a proof of non-implication.
124 known results re-provedCHECKED
Previously known statements proved again using this paper’s results, with all assumptions made explicit. Equivalent statements are counted once.
9 major results re-provedEVALUATED
Of the known results re-proved, these are judged important under the stated assessment criteria. Readers can inspect and challenge that judgment.
Theorem-level citations TRACEABLE
Each proof dependency links to the precise result it uses. Historical priority, inspiration, and attribution remain additional parts of the record.
Submission-time novelty remains attached to the original corpus snapshot. New uses and consequences update the live record.
Three fictional panel members · rubric 01
The rating is debatable. The reasoning is visible.
Illustrative arithmetic mean: 87.43 Panel range: 86.49–88.70 Two decimal places are a display choice, not a claim of certainty.
Panel member A88.70
Strength: unification
The strongest contribution is a common argument that recovers results previously treated separately.
Panel member B86.49
Reservation: scope
Its significance depends on how restrictive the hypotheses are and which important cases remain outside them.
Panel member C87.11
Strength: a reusable method
The framework may matter more than the number of new statements. Its explanatory value deserves separate attention.
Shared rubric: originality, significance, conceptual insight, and exposition. Real reviews would disclose model versions, evidence, and revision history.
03 / Two short essays, in pointsRead the strongest version of each.
The case for. The case against.
Research already claims to reward merit. The question is whether making evaluation explicit would help it keep that promise.
+
The case for
Make judgment inspectable.
Ranking is already here.
Journal placement already confers status. Explicit criteria and published reasons would make those judgments easier to inspect and challenge.
Make AI involvement visible.
If AI becomes part of reviewing anyway, its role, evidence, and limitations should be disclosed. Hidden assistance gives readers less to scrutinize.
A review for every submission.
An initial assessment could be available without depending on a willing referee. Its claims and reasons would remain visible for human correction and challenge.
Recognize work that makes other work possible.
Reusable lemmas, clearer proofs, and repairs could receive traceable credit. These contributions need not arrive as headline theorems.
Find connections beyond the familiar names.
Search across verified relationships and equivalent formulations. A useful result could become discoverable even when its author or terminology is unfamiliar.
Let assessment improve.
New applications, corrections, and objections can update a public record. The original assessment and historical priority remain visible.
−
The case against
A number can reshape the work.
Breakthroughs can arrive unfinished.
A strict formalization gate could exclude a profound idea with repairable gaps. Proof status and mathematical promise must remain separate judgments.
People will write for the metric.
If theorem counts or re-proved results bring rewards, authors have an incentive to optimize those counts. More measurable output need not mean better mathematics.
Merit has more than one dimension.
A definition, a counterexample, or a new viewpoint can transform a subject. Formal correctness and graph position do not settle what is important.
The panel could become a new gatekeeper.
Who chooses its models, rubric, and training examples? A visible score can still concentrate power and favor familiar mathematical tastes.
Dependency is an incomplete account of credit.
The shortest proof may bypass the work that supplied the central idea. Formal citations alone can erase intellectual history and the people who built it.
Precision can harden into authority.
An 87.43 may look more objective than the judgment behind it. If used for hiring or funding, a fallible assessment could become difficult to escape.
The strongest stress test
An unfinished proof can still change mathematics.
A system that confuses “not yet certified” with “not worth reading” would miss the point.
In their notes on Perelman’s papers, Kleiner and Lott describe incomplete arguments and incorrect statements, while explaining that the issues could be corrected using Perelman’s own methods.[1]
The design response: allow a paper to remain published, show exactly which obligations are unresolved, keep an assessment of its ideas, and credit those who supply the missing arguments.
04 / Decisions still openThe rules deserve as much scrutiny as the papers.
What should we refuse to automate?
Who decides which results are major? +
Communities could publish competing rubrics, reference sets, and assessments. A “major result” label should reveal whose judgment it expresses and how a reader can challenge it.
Should there be one overall score? +
A single score makes comparison easy and creates an obvious target to optimize. The proposal includes a panel score out of 100.00; whether it should dominate discovery is a separate, unresolved choice.
How do we reward repairs and reviewing? +
Record who formalized, corrected, generalized, or clarified each result. Traceable credit is a start; recognition, compensation, and career incentives still require institutional choices.
Can an author contest the machine? +
A workable system needs statement corrections, appeals, alternative formalizations, and independent reviews. Assessment changes should carry reasons and a visible history.
What survives when the panel changes? +
Preserve the paper, formal artifact, corpus snapshot, model and rubric versions, and prior reviews. A new panel may change its opinion without rewriting the historical record.