Guidelines
These rules govern how submissions are handled and how the status of a problem changes. They ask nothing of contributors beyond the short format described below. They are provisional: the moderators revise them whenever maintaining the database raises situations they did not anticipate, and in light of the one-year assessment.
I. Participating
Anyone can read the database. To contribute, create an account under your own name, with the email address of your university or research institute or with GitHub, and send a submission: a comment, a result on an existing problem, a reference, a suggested edit, or a new problem. See how to contribute.
You can also write to the moderators (listed on the About page, or at rafael.wagner@uni-ulm.de), for instance to point out a preprint that addresses a listed problem. Results posted elsewhere are monitored by the moderators and added to the problem’s entry.
II. What every submission records
Every result submitted to a problem records, in a fixed format:
- a title;
- the type of material: a claimed solution, a partial or negative result, a reformulation, computational evidence, an independent verification, an exposition, a challenge, and so on;
- the extent to which AI tools were used: none; partially, when the material was produced over several human–AI iterations; or fully, when it is essentially the output of a single prompt;
- the name of at least one human who takes responsibility for the material.
Submissions need not be complete or rigorous. A promising but unpolished argument, an incomplete proof or a partially verified computation is welcome; say in its summary that it is provisional, and how far you understand it.
III. Moderation
Every submission is reviewed by a moderator before it appears, and each decision is logged. Routing every addition through the moderators, instead of opening the database to direct editing, keeps each contribution tied to a named human and excludes automated or anonymous mass submissions.
The check is deliberately light: is the submission relevant to the entry, in the fixed format, and backed by a named human? Moderators filter out only content that is unrelated, predatory or otherwise unethical. They do not judge the correctness of a claimed solution, or the quality of its presentation, when accepting it. If a submission is not accepted, the moderator explains why, and you can follow its status under My submissions.
Moderators may judge the correctness of existing work as independent researchers, in order to change the status of a problem, following the rules below.
IV. The status of a problem
Every problem carries a live status, open, under revision or solved, and every change of status is a dated entry in the problem’s history. A problem enters the database as open.
From open to under revision. When a solution is claimed, whether submitted to the entry or reported to the moderators as a preprint, the problem is placed under revision. Researchers can then submit revisions of, comments on and verifications of the claimed solution; the moderators may comment as well. Submissions at this stage should put human understanding first: commentary on the existing results, independent solutions, verification, replication, or polished, clearly written accounts.
A solution must resolve the problem as stated, by the criterion for a solution fixed in its statement. A proof of a weakened or otherwise modified statement is a partial advance, not a solution: it is recorded as such, the original problem stays open, and the modified statement can be proposed as a new problem.
From under revision to solved. If the solution has been published in a peer-reviewed journal, the problem is marked solved once the moderators have confirmed that the published work settles the problem as stated. Otherwise, it is marked solved only when all of the following hold:
- A human-backed account of the solution exists: at least one human contributor claims full understanding of it, and it is available in a preprint repository that provides a digital object identifier (DOI). Claimed solutions that are entirely AI-generated, or whose contributors do not claim to understand them fully, do not solve a problem; they keep it under revision.
- At least three skilled researchers who are independent of the solution’s contributors and declare no conflict of interest (moderators, for example) have verified the solution, declaring that they understand it and agree with its claim. They may do so by submitting a verification to the entry, by writing to the moderators, or by sharing further material, such as an explanatory manuscript. A machine-checked formalization strengthens this verification but does not replace it.
- There is no current dispute of the solution. An active dispute must either be resolved within the community until its authors withdraw it, or be marked as resolved by a moderator when there is overwhelming evidence that it is.
From solved to under revision. A solved problem is reopened only on a substantiated challenge: a written submission to the entry, backed by a researcher who claims to fully understand it, that identifies a gap or error in the account of the solution, reports a failed replication of a computation or construction, or points to a published result that contradicts the claim. Any researcher may submit a challenge. The moderators reopen the problem once a researcher independent of both the challenger and the solution’s contributors, possibly another moderator, seconds it.
From under revision to open. A problem under revision returns to open in two cases. Its contributors may withdraw the claimed solution, for instance after the open discussion around it (write to the moderators to do so); the material stays attached to the entry as a provisional or negative result, with the credit this implies. Or the revision may show that the statement of the problem itself was significantly flawed: ambiguous, ill-posed, or dependent on an unstated assumption that lacks proof. The moderators then return the problem to open with a revised statement, keep the earlier version and the submissions attached, and credit the researcher who identified the flaw, since clarifying a question is itself a valuable contribution.
Once a solution has been published, the entry stays open to further updates: corrections, simpler proofs, generalizations and follow-ups.
V. Credit
The database records who posed each problem, who advanced it, who verified a solution, who found a flaw and who made a result readable, so that every contributor is credited directly and visibly. Posing a well-motivated question earns credit as much as answering one: cite the problem, and the people who posed it, when you work on it. Every problem page has a Cite this problem button. Taking up a problem posed here without acknowledging its posers is a breach of academic norms.
Everything citable has a permanent link of the form /contribution/<key> that keeps working when pages move: 12 is Problem 12, 12v3 its statement and text as they stood at version 3, 12.3 a contribution to it, and blog.7 a blog post. Problems and blog posts get their key when they are accepted, as do contributions other than references; comments get none. A key is never reused: if what it points to is removed, the link says so. Every accepted change to a problem’s title, statement, text or references starts a new version, except minor corrections such as typos, which moderators apply to the current version.
Improving a problem’s text is a contribution too. A suggested edit names the people to credit and how much AI was used, like a result; once accepted, the new version is credited to them and linked from the problem’s history. Changes the moderators make themselves are credited to the editors, unless they name someone. Each problem lists its contributors: everyone who posed or proposed it, edited its text, or added a result or a reference. Comments are discussion and are not counted as contributions.
Each result can be discussed on its own page, and signed-in readers can vote on it: up if it is useful, down if it is not. Votes are a signal for readers, not a verdict: a claimed solution counts as verified only through the verifications recorded with it. Nobody votes on their own work, and suggested references are not voted on.
Within the database, credit for solving a problem goes primarily to those who understand and verify its solution, and to those who make it understandable. Being the first to submit an AI-generated claimed solution therefore carries only moderate weight, and prompting an AI does not entitle anyone to sole credit for the eventual solution: an AI-generated purported solution can meaningfully accelerate a research direction, but it does not, on its own, solve an open problem. Partial results that are shared here are credited too, even when the final solution is reached by someone else.
Progress on the problems will be documented in progress reports. Everyone who contributed to the discussion or advancement of a problem will be invited to collaborate on, and share authorship of, the report that covers it. The researchers involved in an entry can also decide together whether a result merits a proper manuscript.
VI. The problem list and its cycles
A problem is added to the list only if it meets the four curation criteria: clarity, relevance, significance and feasibility. Proposed problems are checked against these criteria by other researchers and the moderators, not heavily refereed.
For the first cycle, OpenContext I, the list of problems is fixed in advance, to provide a controlled setting for the case study. New proposals are welcome and are kept for consideration in the next cycle; the moderators may admit new problems earlier, depending on the engagement with, and the results of, this first cycle. A problem that becomes obsolete or no longer relevant may be removed.
The long-term goals that organize the list are provisional too. Objections to a goal itself, the assumptions or motivations it rests on and what the goal would lose if the objection held, are welcome: send them to the moderators, who will collect them as starting material for the next cycle.
VII. Best practices
Prompt cautiously, be honest about how far you understand what you submit, avoid single-prompt outputs and duplication, be polite, and recognize the work you build on. The best practices spell these out.
VIII. The one-year assessment
OpenContext I is a test of a hypothesis: that this kind of infrastructure improves the attribution of credit, broadens participation and accelerates progress. One year after launch, the project will be assessed publicly in a community-facing report, against criteria fixed in advance. It will report how many problems were advanced or solved, how quickly, and how much of each advance came from humans, from AI or from their collaboration (as recorded in each submission’s AI-use field); how many researchers, from how many institutions, took part; and whether work that resolves listed problems cites the database and the people who posed them. The dated status changes behind these figures are shown on the progress page.
The assessment will state which parts of the design are open to revision, among them the curation criteria, the status rules above, the moderation load, the size of the list and the credit conventions, and whether to extend the project to further subfields. Negative outcomes will be reported with the same care as positive ones.