Proposer metadata: classify OEIS conjecture proposers by occupation

[tadamcz]

Classifies the 123 unique conjecture proposers in conjecture_provenance.jsonl by occupation, so we can report how the benchmark's conjectures split between professional mathematicians and amateurs.

Approach

Every registered OEIS contributor has a self-written wiki user page, fetchable as raw wikitext from https://oeis.org/wiki/User:<Name>?action=raw (needs a browser user-agent; the default gets 403'd). Only 9 of 123 proposers have no page at all, so this gets most of the way on its own.

scripts/extract_proposers.py fetches those pages, caches them, and asks gpt-5.5 (strict structured output) to classify each proposer into academic_mathematician / academic_other / industry_professional / student / amateur / unknown, plus occupation, affiliation, a verbatim evidence quote, evidence source, and confidence. Resumable; parallel fetch and classification.

Data files

Three layers, deliberately kept separate so judgements stay auditable rather than overwritten in place:

• apn/data/oeis/raw/proposer_pages.jsonl — immutable snapshot of each wiki page's wikitext, so a classification's input survives later edits to the wiki.
• apn/data/oeis/proposer_metadata.jsonl — the wiki-page classification, 123 rows.
• apn/data/oeis/proposer_metadata_web.jsonl — a web-search pass over the 15 highest-weight unknowns (they were 75% of the unknown mass; Peter Bala alone was 36%). Each row carries the verdict, the facts found, source_urls, and method/model/date.

Consumers merge with one rule: a web row overrides a stage-1 unknown, nothing else. That drops unknown from 32% to 7% of conjectures.

Results

Conjecture-weighted over all 492 conjectures, fractional credit for joint proposals:

category	conjectures	share
academic mathematician	241.0	49%
amateur	131.0	27%
industry professional	60.0	12%
unknown	32.5	7%
academic, other field	18.5	4%
student	5.0	1%
no proposer	4.0	1%

The academic mass is highly concentrated: Zhi-Wei Sun accounts for 180 of the 241 academic-proposed conjectures.

Caveats

• Three "amateur" calls (Ordowski, Sigrist, Zschorn) are low-confidence and rest on absence of any professional trace rather than positive identification. If these numbers go in the paper, consider folding low-confidence into unknown; that variant reads 46% academic-math / 21% amateur / 13% unknown.
• source_urls are agent-reported and not archived, so they're exposed to link rot. Snapshotting the cited pages the way we snapshot wiki pages would fix that if we want the trail hardened.
• Two rows in proposer_metadata.jsonl classified as academic_mathematician from wiki text alone look shaky to me and were not web-verified: Klaus Brockhaus and Jonathan Vos Post (6 conjectures between them, so no aggregate moves).