Docs / Seed collection
The seed collection
PubPhys starts from a seed collection of 1,761 problems under 360 topics in 18 fields, compiled for PubPhys on 1 and 2 October 2026 by AI research agents. The problems have not yet been reviewed one by one by a person: each seed problem and topic is labelled "Compiled by AI, not yet reviewed by an editor" until an editor's review record exists. It is a starting point: every problem can be refined, split or contested.
How it was prepared
- Duplicates were merged: 58 problems and 5 topics that asked the same question were folded into one copy.
- Every topic's review reference was checked against Crossref or arXiv; 3 were corrected.
- Status notes that mention work from 2024 on are marked unverified (742 notes): they were partly written without a live literature search.
- What was checked by people: the automated audit of every review reference above, and a sample of the formula conversions below. Nothing else was reviewed by hand.
- The signed seed records leave
assisted_byempty: their provenance is the import manifest below, which says how they were made. A later revision of a seed problem names the AI systems it used: a revision 2 record lists the AI model that repaired its notation, which the problem page shows as "AI systems used to prepare this seed revision, as listed in its record" (no author declared it), and an admin's edit lists what that admin declared. - Formulas written in plain text were converted to typeset notation by AI. Each formula was checked automatically: the text around it is unchanged, it covers whole words of the original with the same symbols in the same order, and it typesets. Where only some of the proposed formulas passed, those were kept and the rest left as plain text. A random sample of the conversions was reviewed by eye.
Revision 2: notation repaired
Seed revision 2 repairs the notation of the seed text: formulas left partly converted, numbers such as 1e-19
inside formulas, words used as operators, plain chemical formulas from a reviewed list, and question marks in titles. A
fixed rule table made most edits; an AI model repaired partial conversions, each edit checked by a validator that keeps the
numbers, symbols and the text around them. Every edited field is listed in the batch's items file, saying whether a rule,
the AI or a person changed it. Fields whose repair failed kept their text and are fixed by hand later.
Nothing was overwritten. Revision 1 of every problem and topic stays in the log unchanged; revision 2 is a new signed record that names revision 1 and the batch's import manifest as parents, and the platform's review of it says it was accepted with the batch. The label "Compiled by AI, not yet reviewed by an editor" stays: a platform review is not a person's review.
The batch approval. One person, the batch approver, signed with his ORCID iD one review record accepting the batch's manifest. The manifest's file names the SHA-256 of the items file and of the review file, so the approval is bound to the exact data and to the decisions taken on the review page. Before signing he had the counts, those hashes, the review lists (a random sample of rule edits, every chemical formula rewrite, every AI-assisted field, the doubtful titles, the failure list) with the text before and after, and both files. Decisions on the review page were taken by people, some by agreeing with an AI suggestion; the review file records which, per item. The approval approves the batch, not each problem: it does not make any problem reviewed by an editor.
A bundle proves one record. The verifier checks a record's own rules and signatures, not what its
problem_ref or parents point to. To check the chain, verify the revision 1 bundle, the revision 2 bundle and the
review bundles separately, and resolve the references on this site or from an archive of the log.
While a batch is under way
A seed revision batch writes over the revisions it was prepared against. So, while a batch is prepared, approved, or written and not yet anchored in Bitcoin, changes to seed problems and topics wait: editors' revisions of them, the decisions on those revisions, moving a problem to another topic, marking a status note verified, granting an appeal of a decision on one, and removing a revision for abuse or spam. The administrator's own seed edits start once the revision 2 batch is anchored, not merely written. A person's own problems and topics never wait.
Removals for copyright or personal data never wait. A batch they touch before it is written is set aside and prepared again without that item. When a seed revision must be withheld for copyright or personal data, its text is replaced and the problem shows its previous revision; this move is recorded in the admin log, not as a signed record. If a result answers the revision's question and the revision before it asks another, the revision stays displayed with its text withheld until the result is decided or removed.
- Import manifest: ce59fb6cd812fc05…
- Batch approval by Victor Galitski: c4417ea89cbb0b93… (bundle)
- The batch's files, the exact bytes the manifest and the approval are bound to: items.json revision2-review.json report.json manifest-index.json
Later batches
A later batch changes seed problems and topics the same way: under its own name, with its own files, approved by the batch approver with ORCID before it is written, one batch at a time. Each item names the revision it was written from, and the batch is refused if that revision has changed since or if a text it changes breaks the rules of an edit. A citation moved into a problem's references is copied by the batch's code from the text it was moved out of, not written by AI, and listed for review on its own; one that names only an author and a year is read one by one. Once written, a batch stays in the log: turning the setting off stops new batches, it does not take a written one back. Each batch written is described below with its files.
Timestamp
The import manifest, which lists every topic and problem with a hash of its content, is timestamped in Bitcoin through OpenTimestamps. Anyone can check that the collection existed in this exact form no later than the anchoring block.
- Manifest: 362bc6b25f5a841c… Bitcoin #969,704
- Item list: 614d670cecf03d64… Bitcoin #969,704
Imported 2026-10-03 07:17 UTC.
AI summaries
Below the signed statement of a seed problem, and below the signed summary of a seed topic, a box headed "AI summary" may give the same text rewritten for specialists: shorter, in the register of a research paper in the subfield, with every symbol named where it first appears (once enabled). It is shown in Expert mode only.
- What they are: the summaries were written by an AI system from the signed texts of the problem or topic alone. There was no second reader and no check against the literature. Nothing in a summary is meant to go beyond its signed text.
- They are not signed. The signed text above the box is the record; a summary is not part of any record, its bundle or its Bitcoin timestamp, and it is not in the API, the search, the MCP tools or the sitemap.
- "Checked by an editor" means that the chief editor read the summary against the signed text shown on the page and kept, corrected or released it. Moving a summary onto a newer signed text (a rebase) and releasing one that was held count as such a check. Until then the box says that no editor has checked it.
- Some summaries are held back and not shown: those whose writer noted a dropped detail or an undefined symbol, whose numbers or units do not all appear in the signed texts, or whose formulas break the rules of the seed's notation. The chief editor releases them one by one.
- When a problem's signed statement or what would settle it changes (a topic's summary, for a topic), its AI summary is hidden until the chief editor has checked it against the new text.
- A summary that misstates its signed text can be reported through the contact form, with the address of the page.
Starting numbers for the ranked list
Every topic and problem of the seed collection gets a starting number from 0 to 100, the editors' number with which it enters the ranked lists before members vote (how ranking works). Starting numbers were drafted by AI agents outside the site, for 360 topics and 1,761 problems. The chief editor reviews them before ranked lists are turned on and can change any of them; each change is recorded. A problem or topic without a number starts at 45.
- One scale for all fields. A number answers one question: if this problem were settled, how much would science, technology and research programs change over the following decades. A topic's number is the consequence of settling the topic's open questions as a whole, on the same scale, and stands at or above the number of its highest problem.
- Six scorers, each with whole fields, read every problem's record and gave every problem and topic a number with a one-sentence reason, and a flag where one applied (for example a problem that repeats another).
- Independently, a calibrator scored all 360 topics, and an adversary chose its own 150 most important problems from the titles alone. Then 54 challenges were raised against the numbers (a proposed change of at least 8 points, a crossing of 70, 85 or 95, or a flag) and answered in 87 revisions.
- The two independent sets of topic numbers agreed with a rank correlation of 0.935 before any exchange, and 145 of the adversary's 150 problems stand within the first 300 of the list.
- Each problem has one number from one scorer, so a number is good to several points, and ranks deep in the list are coarse. Many problems share a number. The reasons and flags were not imported; flagged problems are for the editors to revise through the usual forms.