What was built

Everything is under publish/ plus scripts/publish-notes.sh; nothing committed.

The brief said eleven notes; there are ten (plus the README), so the drafts cover nine.

Reports published (44): every report named in the notes’ Artefacts sections (abbreviations expanded), in the real manifest, and in the notes README, and all of 2026-10-01-*.

How to run the three scripts

From the private repository root:

scripts/publish-notes.sh --check          # stale/absent targets; exit 1 if any
scripts/publish-notes.sh                  # copy; refuses (copies nothing) if a source is missing
.venv/bin/python publish/scripts/validate.py
.venv/bin/python publish/scripts/build_site.py

Variants: validate.py --dir manifests/_drafts --also manifests; build_site.py --manifests manifests/_test (or add --manifests manifests/_drafts for a preview; docs/ is rewritten in full each run); --base-url; NOTES_CSL. PyYAML and jsonschema were installed into the venv. A build takes about 8 s.

Current results: sync 70 listed, 0 stale. validate.py: 1 manifest, 32 claims, 0 errors, 3 warnings (the real manifest’s periodicity cites curl-reduction/thm-p, lemma-a, lemma-l; with --also manifests/_drafts these resolve and there are no warnings, because the curl-reduction draft uses those exact ids). Drafts: 9 manifests, 76 claims, 0 errors, 0 warnings. build_site.py: 0 errors and one notice: reports/2026-09-04-certificates-sat.md has a --- rule that pandoc reads as a YAML block, so the builder re-renders that one file with metadata blocks off and says so. Negative tests: a tampered sha256 and an invalid status both give exit 1. Every JSON-LD block (87) parses, every relative link in docs/ resolves, and the inline JavaScript passes node --check. I could not open the pages in a browser (no Chrome in the sandbox), so the graph’s layout and the MathJax rendering have not been seen.

The site, page by page

• index.html: a paragraph on the project for a reader who knows nothing of it; counts (claims, manifests, notes without a manifest, claims read by a person: 0 of 32); links to graph, notes, reports, certificates, llms.txt, llms-full.txt, schema, repository; an Open problems box (status open or conjecture, plus each manifest’s open_problems: now thm-h and thm-u) with statements; the claims table (id, label, kind, status badge with the gloss as tooltip, scope, note), filterable by status, note and free text, and by ?status= / ?note= in the URL; the status vocabulary with glosses, parsed from the schema so it has one source.
• graph.html: d3 7.8.5 from cdnjs; nodes coloured by status (colours from CSS variables, so dark mode follows), arrows from a claim to the claims that use it, laid out in levels by longest dependency chain (foundations at the top), dashed arrows for supersedes, grey nodes for cited claims of notes with no published manifest yet; pan, zoom, drag, hover for full label, click to open the claim page; a legend; a <noscript> note and a collapsible plain list of the same edges.
• claims/<note>/<claim>.html (32): label, the statement in a box edged in the status colour (MathJax 3.2.2 from cdnjs; statements, details and notes go through pandoc so Markdown and TeX render), then Identifier, Kind, Status with gloss and detail, Scope and hypotheses, Depends on and Used by (computed), Supersedes / Superseded by, Artefacts (reports linked to their pages, ledgers to the copies on the site, private: paths shown as unpublished, size, sha256, regenerate command, note), Formal, Provenance (written by, checked by, “Read line by line by a person: no”, date, log), Notes, Source (note and manifest JSON). JSON-LD: a CreativeWork with identifier, genre (kind), creativeWorkStatus, text, isBasedOn (its dependencies), isPartOf the note as a ScholarlyArticle, subjectOf the manifest.
• notes/index.html and notes/<slug>.html (10): the table of notes (title, the “contribution” cell of the notes README, date, manifest, PDF). Each note page is pandoc (--standalone, MathJax, --toc in a collapsible box, --citeproc with notes/refs.json and the New Oxford CSL), headed by links (PDF, manifest JSON and YAML, Markdown source), the README’s certification cell, and the list of the note’s claims with badges (or “no published claims manifest yet”). Code spans docs/reports/X.md become links when X is published. JSON-LD ScholarlyArticle with hasPart (one CreativeWork per claim) and the PDF as encoding. PDFs are copied into docs/notes/.
• reports/index.html and reports/<name>.html (44): newest first, with title, file name and the claims that cite each report; each page notes it is published as written and lists the citing claims. JSON-LD Report.
• certificates/index.html: per ledger, a table by statement, cell and tag: queries, UNSAT, SAT, DRAT verified by drat-trim, proofs kept, proof size; the ledger’s own sha256; the citing claims; the raw .jsonl and _sha256_ files are copied alongside. Totals now: h28x 114 queries, 114 DRAT-verified; h74x 405 queries, 156 DRAT-verified (the 155 of the campaign plus one main-loop recheck at (7,7,7)). The statement letter is the ledger’s stmt, or for h74x the letter in the query name.
• llms.txt (summary blockquote, the status vocabulary, a caution to read status_detail and scope, the manifests, the full-text file, the schema, the notes with one line each from the README table and their PDFs, optional links); llms-full.txt (every manifest as JSON with a header per note); manifests/<slug>.json; schema/manifest.schema.json (copied so the schema’s own $id URL resolves on Pages); sitemap.xml (91 URLs).
• style.css: one file, light and dark, phone-width layout, no framework.

Draft manifests (for review; not on the site)

Written by three Opus sub-agents from each note and the README; validated here; curl-reduction/thm-p spot-read.

What the drafters flagged for the main loop:

1. docs/notes/README.md is inconsistent with itself: “Where the older notes are superseded” says periodicity holds on cubic sides one to six, open at seven; the table and the 1 October paragraph say side eight. The drafts quote “one to six” where they cite that paragraph. The paragraph needs updating (not mine to edit).
2. The README says Theorem P was “corrected to require every extent at least two”; curl-reduction’s Theorem P already has that hypothesis. The correction was to the periodicity chain as theorem-h-orientation’s first draft stated it.
3. Two incompatible statements are both called “Theorem H” (curl-reduction: six directions imply even curl; theorem-h-orientation: an axis empty or one-way). No supersedes link was drawn; consider one, or a note on theorem-h-orientation/thm-h.
4. sector-logconcavity: the abstract says all 24 real-root verdicts are certified in Isabelle; the body certifies 13 in full and only a root-count deficit for the other 11, whose “not real-rooted” rests on a Python squarefreeness computation. The draft splits it (certified / computed); the README repeats the abstract’s wording.
5. CPOG model counts were given certified, though the schema’s gloss names only DRAT, LRAT, Lean and Isabelle.
6. checked_by in the drafts lists second enumerators and decision procedures named in the notes, which are not independent agents.
7. The method note’s status words are not the schema’s; some artefacts are directories or globs.

What I would change in the schema

The schema is unchanged. Suggestions, most useful first:

1. certified should name the checkers it admits, CPOG included (or say CPOG counts are computed). Two drafts depend on the answer.
2. Separate independent readers from cross-checks. checked_by is defined as “independent checkers that read it with their own programs”, but the older notes’ evidence is second methods in the same codebase. Add an optional cross_checks list (second enumerator, three decision procedures, drat-trim) and keep checked_by for independent readers. Today drat-trim sits in checked_by in the real manifest.
3. A machine-readable cell list. scope.cells is free text; an optional cells_list (array of [a,b,c], plus a family string for infinite families) would let a reader or a tool ask “does this hold on (6,6,9)?”.
4. Artefact patterns. Allow kind: directory or a glob: true flag, since notes cite directories and globs.
5. A status for findings about practice (the method note), or leave that note without a manifest.
6. definition with status proved (the real manifest’s model) reads oddly; a defined status, or allowing status to be omitted for definitions, would be clearer. # What is left

• Create lyndondrake/ice-mixing-notes on GitHub (public), push publish/ as its root (the sync is file-level; a git subtree split --prefix=publish or a fresh clone with rsync of publish/ are both possible), and set Pages to “Deploy from a branch”, main, folder /docs.
• Decide whether the private repository tracks the generated publish/docs/ or ignores it.
• Zenodo: enable the GitHub integration for the public repository, cut a release (tag), let Zenodo mint the DOI, then record it in CHANGELOG.md, each manifest’s doi, and CITATION.cff (doi:, version:, date-released:). Add .zenodo.json if the deposit metadata should differ from CITATION.cff.
• Review the nine drafts against the log and promote them to manifests/.
• Seven older reports contain absolute paths to home directories (~/..., ~/...): 2026-08-11-stage1-falsifiers, the five 2026-09-04-* in the list, 2026-09-17-completeness. Harmless but personal; publish as written or redact in the copies.
• notes/README.md is published (as the brief asks) but contains private build instructions (docbuild.sh, Spark paths); the site does not render it.

Rules kept or broken

• Kept: wrote only under publish/ and scripts/publish-notes.sh; did not modify docs/, formal/, experiments/, LOG.md, HANDOVER.md, the root README or any existing script; Python from the project venv; temporary files in .tmp/; no /tmp; no commits; no network fetches (cdnjs URLs are only written into the HTML); UK English and no em-dashes in what I wrote (checked by grep); no statuses, checkers or hashes invented: the site renders what the manifests say, the certificate counts are computed from the ledgers, and the drafts were held to the notes and the README.
• Departures: the brief said pandoc --mathjax; pandoc 3.11 deprecates the flag, so the builder uses the equivalent --math-method=mathjax:URL. The first draft-manifest sub-agent was dispatched with an unfilled placeholder for its notes; it wrote nothing, and I sent it the three slugs. I edited one drafter’s output (method.yaml: a directory artefact reports/ became private:docs/reports/) so the drafts validate. The test fixture in manifests/_test/ is in the public tree; delete it before the first push if unwanted.