AI at PubPhys

AI helps with a few things and decides nothing on its own: it proposes the fields of a new problem from the author's file; it screens a signed problem, a later revision of one, a proposed topic, a signed result and a signed contest (with the text of their source files), and comments, so that a clearly fine one need not wait for an editor; and after publication it checks the problem for gaps and possible duplicates, as a private report for the author and the field's editors. Everything AI produces is labelled, the author confirms every field before signing, and anything doubtful goes to a person.

Only with consent

When uploading, the author chooses whether AI may read the file and the fields made from it (off by default). The choice is part of the upload's signed record, logged before any AI call. Without consent nothing the author wrote is sent: the author fills in the fields and an editor screens the problem. A published problem may still be compared with others deterministically, by its words and without AI.

What is sent, and to whom

• Recipient: an external AI provider, through its API.
• Formalization: the uploaded file (a PDF up to 30 pages as is, otherwise its text) or the typed text, and the titles of the field's accepted topics that may be sent (those of the seed, and those whose proposer consented to it). Task formalization/2; when topics are being proposed and none of them fits, it may draft a new topic (title, summary, why it matters), which the author edits before signing it with the problem (proposing a topic).
• Screening of a topic: only under the consent asked when the topic is signed (version t1), given for that revision: its title, summary, why it matters and review citation, with the titles of the field's accepted topics that may be sent, as above; not the proposer's name or iD. Task topic_screen/1; it looks for a coherent area of physics that is not a single question, a duplicate or spam, never decides acceptance, and doubt sends the topic to an editor. The same consent lets later formalizations and topic screenings send the topic's title once it is accepted. The consent text: AI may read this topic as you sign it. PubPhys sends its title, summary, why it matters and review citation, with the titles of the field's accepted topics, and nothing else, to an external AI provider, which checks before publication that it is a topic of physics, not too narrow, not a duplicate of an existing topic, and not spam. The screening does not decide whether the topic is accepted; editors do that. Once an editor accepts the topic, its title may be sent, as one of the field's topic titles, in the screening of other topics and the formalization of new problems in the field. The provider does not use API data for training and is asked not to store it (store: false); under its standard terms it may keep API data for up to 30 days to detect abuse. Without this consent nothing you wrote is sent: an editor of the field screens the topic before it is published, and its title is never sent.
• Screening: the four signed text fields (title, plain words, precise question, what would settle it). For a problem with parent problems, also each declared relation and note, and each parent's title and the first 400 characters of its precise question when that parent may be sent as an excerpt (below); a doubt about a relation sends it to an editor. A variant started from a problem page has no file and no formalization; its consent covers the screening and the check after publication.
• Check after publication: the signed fields of the published problem, the field's name and the topic's title when it may be sent (as for formalization), with the titles and the first 400 characters of the precise questions of up to five published problems that look similar by their words, its declared parents first when its author's consent names parent problems (otherwise a parent is sent only if it is among the similar ones, and is not marked as a parent). Under the problem consent of version 3, a declared parent that is similar by its words may thus be sent as one of the up to five similar problems that text covers, ranked by similarity alone and without its relation, its note or any mention that it is a parent. A problem is sent as such an excerpt only if its author consented to it (the consent text since October 2026 says so) or it comes from the seed. The check may be repeated while the problem is public. Its report goes to the author and the field's editors, is never public, and acts on nothing.
• Screening of a result: only with the author's consent for that very result, given before signing and bound to its fingerprint: its kind, its text, its references, and the text of its source files. Not the problem, not the author's name or iD. The screening asks whether it is a serious, intelligible piece of physics work; it does not judge correctness, and doubt sends it to an editor (the stages of a result).
• Screening of a contest: only with the filer's consent for that very contest, given before signing and bound to its fingerprint: its grounds, the passage it quotes, its text, its references, and the text of its source files. If the text it quotes is withheld before the screening reads it, nothing is sent and an editor looks at it first. Not what it contests, not the problem, not the filer's name or iD. The screening asks whether it is a serious, intelligible objection in physics; it does not judge whether the contest is right, and doubt sends it to an editor (the stages of a contest).
• Screening of a comment: only if its author allowed AI screening of comments on their account (asked once; changeable at any time): the comment's text at each version, and nothing else (not the author, the problem, the result, or the comment it answers). Task comment_screen/1; it looks for spam, abuse, personal data and unintelligible text, not for correctness, and doubt sends the comment to an editor (comments). Without consent, editors read each comment first.
• Screening of a later revision of a problem: only under the consent asked in "Edit this problem" (version r1), given for that revision: its signed fields, the changed notes (references, status note, when it was posed) with the four fields of the question, and nothing else. Stage 1 screens it before publication as it screens a new problem; a core change is checked again after publication (stage 2), with the titles and excerpts of similar problems, a cosmetic one is not. An editor's revision of a seed problem is the editor's text, under the editor's own consent. Without consent an editor screens it first (revisions). The consent text: AI may read the fields of this revision as you sign them. PubPhys sends them, and nothing else, to an external AI provider to screen them before publication. If the revision changes the question itself (a core change), its signed fields are checked again after publication, together with the titles and short excerpts of up to five published problems that look similar, to report gaps and possible duplicates to you and the field's editors; these checks may be repeated while the revision is on display. The provider does not use API data for training and is asked not to store it (store: false); under its standard terms it may keep API data for up to 30 days to detect abuse. Without this consent nothing you wrote is sent: an editor of the field screens the revision before it is published.
• Source files of a result or a contest (at most 3): their text only, read on PubPhys's server (a PDF's text by poppler, out of process; never the PDF itself), at most 40,000 characters per file, each between markers that no file can contain, as data. A file that cannot be read as text (a scanned PDF, say), has no text, or is longer (or a PDF of more than 30 pages) is not sent: the result or contest is held for an editor instead (hold reasons file unreadable, file too long). A file is sent only under the consent given for the item it belongs to, which must be the choice made when the file was uploaded.
• Nothing else: no name, e-mail or account data. Requests ask the provider not to store them (store: false); the provider does not train on API data, and under its standard terms may keep it for up to 30 days to detect abuse.

Declared assistance

When the author started from an AI proposal, the signed record says so: both origin.assisted_by and content.assisted_by are ["pubphys:ai-formalization/1"] (otherwise both are empty). PubPhys enforces this, so the declaration cannot be dropped, and anyone can check it in the record.

A result declares the AI systems its author used, up to 5 names, in both places of its signed record; PubPhys requires the two to be identical and shows them on the result's page.

Manipulation

A file can try to steer AI. Files are treated as data, never as instructions; a problem whose text seems to address the AI is held for an editor, and AI output is shown with the warning that it may be manipulated by the submission. The same holds for results: a result whose text seems to address the AI is held for an editor of the field (hold reason injection suspected), never published by the screening.

What a check after publication says

Whether the problem can be understood on its own, gives what is needed, is free of contradictions and is neither over- nor under-determined; whether another published problem asks the same question; possibly a field that fits better, an expertise to consult, or a hint that the question may already be answered (unverified, from the model's own knowledge). A report is marked for an editor's look only on a confident answer. In this phase nothing follows from it by itself: editors acknowledge or dismiss it; merging duplicates and asking for changes come later.

Limits

PubPhys keeps a daily spending limit for AI, per UTC day. Screening a signed problem, revision, topic, result or contest may use all of it; the other tasks keep a share free for that screening: AI proposals of a problem's fields and the screening of comments stop while less than 10% of the day's limit would be left, and checks after publication while less than 30% would be left. When it is reached, or when the provider is unavailable, nothing breaks: authors fill in the fields themselves, new problems wait for an editor, and checks after publication wait until the next day. A result is held for an editor of its field at once when AI is switched off or the limit is reached; when the provider fails to answer, the screening is retried for up to six hours and the result is then held (held results).