Agents

Public notebook of finite-handle attacks on open problems: https://github.com/wustep/maths. Stephen Wu is the human author. Models go in the README ledger. CLAUDE.md is a symlink to this runbook.

Problem folders

One folder per problem under problems/<slug>/.

PROBLEM.md              statement and what would count as a new bound
ATTACK.md               chronological attempts
WALKTHROUGH.md          discovery notes, not a cleaned proof
RESEARCH.md             papers, OEIS, failed lookups
compute/                verifier plus certificate
compute/LEAVES.md       live finite-handle queue (open / certified / residue)
compute/q<n>/           one quest
compute/q<n>/CLAIM.md   exact inequality this quest asserts (if any)
lean/                   lemmas for this problem, if any

Keep ATTACK.md chronological. Walkthrough beats: refs/walkthrough-style.md. Cite URLs you opened in RESEARCH.md. lean-toolchain pins Lean 4.32.0. Lemmas live in the problem folder.

Starting a problem: mint with scripts/new-problem.sh <slug>; recipes in .claude/skills/ (new-problem, literature, markdown-latex, compute, writing); fetch a paper with python3 scripts/arxiv_fetch.py <id> (optional --research problems/<slug>/RESEARCH.md); OEIS via scripts/oeis_lookup.py. Then add a Problems-table row. Add a model-ledger row when you run it.

How to work

Read PROBLEM.md and the published record first. arXiv is the record; fetch and replay before trusting a number. Forum numbers (MSE, Reddit, MathOverflow, AlphaXiv) are leads, not citations.

Before attacking, also read compute/LEAVES.md. If it is missing, create it from the finite handles already in PROBLEM.md. Prefer an open leaf. One leaf per quest. If you invent a new finite handle, add a LEAVES row before hunting.

A new bound is a verified finite improvement of a published record. An incomplete search is not a lower bound. Do not invent a status noun beyond these two:

dent — verified finite improvement of a published record. Write the inequality.
residue — incomplete search (holes, SAT UNKNOWN, timeout). Not a lower bound.

LEAVES rows use open | certified | residue | blocked. certified is a dent; residue is the same leftover as above.

A quest is a named campaign, not a result. Its code lives in compute/q<n>/. ATTACK.md may say q3. Do not write "quest" on README, explainers, PROBLEM, WALKTHROUGH, or other human pages.

Every quest that asserts a finite outcome keeps compute/q<n>/CLAIM.md: the exact inequality or predicate, the prior record, and what would falsify it. run_all.sh exit 0 means the CLAIM holds. PROBLEM and README quote the CLAIM; do not paraphrase a different number. A residue wrap still updates LEAVES to residue and may keep the CLAIM that was attempted.

Lean is optional in problems/<slug>/lean/ for finite algebraic or modular claims. Do not require Lean for SAT, search, or anneals.

Do not commit bulk *.drat / *.cnf / *.g6 (or packed proof shards). Prefer a tiny summary JSON plus a regenerator. Multiple worktrees are fine; do not copy giant cert trees into every campaign. See .gitignore and problems/*/compute/STORAGE.md.

README and explainers use ordinary English; ATTACK, WALKTHROUGH, PROBLEM status lines, and skills may use dent and residue. Independently verify every claimed number. Cite the record you beat, or say you did not beat it.

User-facing prose follows .claude/skills/writing.

What to update

Update the problem-folder files. If status changed, update the README Problems row. When you run something, add a row to the README “Which model ran what” table. Ledger cells name the model (and date), never OpenAI plan/workspace labels such as 20xx or wustep. For a dated campaign, add a folder under notes/ like notes/supergrok-2026-08-17/ (REPORT + prompts; JSONL logs stay local / gitignored).

Anything sent outside the notebook is copied to problems/<slug>/share/<YYYY-MM-DD>/ (the whole problem folder at that send) and then not edited. A git tag on the send is the pin. Later search does not go in share/.

problems/ is current. notes/lists/ catalogs, notes/picks/ ballots, notes/chat/ the transcript. Those folders stay.

Site

web/ is a Next static export. Skip file/dir collisions (fileRouteCollides). Hide CLAUDE.md and vercel.json in EXCLUDED_FILES (web/src/lib/repo.ts). Show AGENTS.md. Production: https://maths.wustep.me, https://maths-wustep.vercel.app. README math on github.com: .claude/skills/markdown-latex.