This documents the Briefing item pattern first explored in the former Week 31 candidate page, together with the current Review roles, so the markup does not need re-deriving per page. That page was removed rather than retained as a collection entry or public example; About Briefings now uses “Attention Is All You Need” to illustrate the format boundary. Every retained markup choice below was verified against both the built site and Obsidian’s own preview—this project’s authoring tools—rather than the site alone; a pattern that only renders correctly in one is treated as broken.
File location and role. A Briefing goes under site/content/briefings/. A Paper Review goes under site/content/reviews/paper/; a Topic Review goes under site/content/reviews/topic/. See Research content storage and routes for the full scheme. The current corpus uses role folders without subject subfolders; field handles subject coverage. Every content folder is also a real page, so add a new subfolder only together with a meaningful index.md, never as an invisible storage convenience. A single Topic Review may separately graduate to its own per-topic subfolder (reviews/topic/{slug}/index.md plus sibling satellite pages for program-specific formal apparatus) once its reconstruction needs it — see Research content storage and routes and DESIGN.md’s Information architecture; this is unrelated to the still-undecided question of subject subfolders addressed above.
Listing frontmatter. A listed Briefing declares collection: briefings and format: Briefing. A listed Review declares collection: reviews and format: Paper Review or format: Topic Review. All three roles also supply date and field, with optional summary. unlisted: true keeps a draft out of ordinary collection indexes and discovery surfaces but does not replace human review or explicit publication approval. npm run preview:drafts may be used locally to show those pages in their matching hub with a Draft label; it does not change their publication state.
Paper Review source metadata. Before the body, use one compact Markdown table with Authors, Publication, and Reviewed version. Publication links the canonical venue record or DOI; Reviewed version identifies the exact manuscript or revision actually checked and states when supplementary material is included. This keeps basic provenance available without turning the opening into source-process narration.
Paper Review body structure. Put the paper-facing account under ## Main text. Use numbered ### headings in reading order: add ### 1. Introduction when the paper’s opening is unheaded, then retain the paper’s own section titles rather than substituting a generic Summary / Methods / Results / Discussion outline. The numbers are Review navigation when the source itself is unnumbered; preserve source numbering when it exists. Put source checking, interpretation, comparisons, and material limits inside the corresponding source section so that scrutiny stays close to the claim. Do not add default What the paper establishes, Limits, or Assessment sections.
An optional ## Supplementary information follows Main text only when source-supplement detail materially helps the Review. It may contain extended derivations, implementation detail, additional tables, solver and cost analysis, sensitivity or convergence checks, and selected appendix material. Follow the source supplement’s headings and order for the parts included, but do not reproduce every appendix merely for completeness. Omit the whole H2 when unused; never move central evidence, conclusions, or important uncertainty out of Main text. ## References follows afterward. site/templates/paper-review.template.md is the repository starting point. Additional editorial sections remain a later decision to make from real completed Reviews.
Scientific terminology in Korean drafts. Keep established English academic terms when translating them would make the sentence less natural or make correspondence with the source harder. For example, use quench, bond dimension, pseudoinverse, and sign problem rather than forcing uncommon Korean equivalents. Define an abbreviation or potentially unfamiliar term at first use, then use it consistently. General connective prose remains Korean; ordinary words such as method, result, and system do not become English merely because they appear in a technical paragraph.
Paper-specific figures. Keep a figure made for one Paper Review beside its owning Markdown file and prefix the asset filename with the page slug, for example carleo2017solving-rbm.htm; do not create an asset-only subfolder under site/content/. Reference it with a relative path so the page and figure move together. Use descriptive accessible text and show the source or adaptation credit inside the figure. Prefer a normal image asset for a fixed illustration; a script-free HTML figure is appropriate when equations and labels must remain legible by reflowing between desktop and mobile. Use .htm rather than .html for this asset type because Quartz reserves .html as a page extension and removes it during slug generation. Site-wide shared images remain under site/quartz/static/.
Grouped summary. For a multi-item roundup, prefer a YAML block-scalar summary: | over a single prose line — a run-on sentence weaving every item together stopped being readable once an edition passed four or five items. Write one field-label line per subject section, each followed by its items as - bullets, in the same field order as the ## sections below:
summary: |
Quantum Information & Computing
- Dimension-free polylog shadow tomography
- QEC syndrome-timing scheduling
Many-body/Materials
- 2D Hubbard strange-metal mechanism
ML/AI
- Optimal agnostic PAC learningThe card renders each field label as a link into that field’s ## section on the item’s own page — the anchor comes from a small FIELD_ANCHORS lookup in collection-index.tsx keyed by the exact field string, not by slugifying the label itself. FIELD_ANCHORS always needs an entry regardless of wording (Quartz’s own heading slugifier output isn’t guessable enough to compute at render time), so it’s checked in explicitly rather than derived. Where the field value reads naturally as a heading, use it verbatim — ## Quantum Information & Computing, ## ML/AI, ## Mathematics — so there’s only one wording to keep in sync. Many-body/Materials is a deliberate exception: its / has no surrounding spaces, and Quartz’s slugifier only inserts a hyphen where a literal space sits next to the stripped character (confirmed against built output — ML/AI → mlai, no separator, because there’s no space either side of the /; Quantum Information & Computing → quantum-information--computing, because there is). A heading of exactly ## Many-body/Materials would therefore slugify to the unreadable many-bodymaterials, so the body heading instead reads ## Many-body / materials physics — natural prose, spaced slash — which slugifies to the readable many-body--materials-physics; FIELD_ANCHORS["Many-body/Materials"] points at that id. Follow this same logic for any future field whose compact taxonomy label wouldn’t read well, or wouldn’t slug well, as a literal heading. Add a new field’s anchor to that map, alongside the citation subject-code table above, before using it in a grouped summary — an unmapped label still renders as plain (non-linked) text rather than failing the build. Each group shows at most three items before collapsing the rest into a linked “+N more”; list every item in the frontmatter regardless (the cap is display-only), so the underlying data stays complete if the cap changes later. A single-line summary (no - bullets) still renders as a plain paragraph, unchanged from before.
Roundup title and date. A multi-item roundup title is the plain publication date, Month D, YYYY, for example August 7, 2026, matching the frontmatter date used for ordering and the card’s date label. (An earlier Week N, YYYY convention, keyed to the ISO week as an editorial snapshot window, is preserved via alias on the items that used it — see 2026-08-06.md, formerly 2026-w32.md — but is no longer used for new items: once a second roundup could land in the same ISO week, a single week number stopped uniquely identifying a page.) A roundup title does not say that every source appeared on that exact day, and does not promise a fixed publication cadence. Subject scope belongs in field tags rather than the title.
Heading. ### P1. Title / ### A1. Title — a one-letter source-type prefix (P Paper, A Announcement) directly on the item’s own number, reset per ## subject section. No separate badge or word: a styled element embedded in a heading leaks into the page Table of Contents as literal HTML source rather than rendered text (Quartz’s TableOfContents transformer calls mdast-util-to-string with its default includeHtml:true), so the type has to live in the heading’s own plain text.
Metadata table (Authors/Venue/Tags). A genuine markdown pipe table, not raw HTML — three rows, Tags included, with Authors itself as the table’s real GFM header:
| **Authors** | Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, Illia Polosukhin |
|---|---|
| **Venue** | [*Advances in Neural Information Processing Systems* 30 (2017)](https://proceedings.neurips.cc/paper_files/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html) |
| **Tags** | [[attention\|Attention]] [[transformer\|Transformer]] [[machine-translation\|Machine translation]] |The Venue cell states what the source is without editorializing on its review status: a plain link ([arXiv:2608.06258](https://arxiv.org/abs/2608.06258), optionally with a year) is enough once the link itself already shows it’s an arXiv/bioRxiv posting — don’t also prefix it with the word “preprint” (redundant) or a parenthetical like “(not yet peer-reviewed)” (editorializing the reader doesn’t need; the venue already tells them). Spell “preprint” out in prose only when the cell has no direct link to the preprint itself.
An earlier version of this pattern used an empty | | | header (GFM requires a header row to exist even though this table’s rows are labels, not columns) and hid it with display: none on the built site. That worked, but Obsidian doesn’t load that CSS, so it showed a harmless-but-pointless blank bar above Authors. Making Authors itself the header instead means Obsidian shows something real there and the site shows nothing extra either way — custom.scss neutralizes the browser’s own default <th> look (bold, wider padding, its own border) so it still matches Venue/Tags below, scoped to a table whose body is exactly two rows so a genuine multi-column table elsewhere (like the one on Research tools and sites) keeps its real bold header untouched.
The Tags cell needs a backslash before each wikilink’s | alias separator ([[slug\|Label]], not [[slug|Label]]) — inside a table cell, an unescaped | would otherwise read as a new column boundary. This is standard GFM table-cell escaping and both this site’s build and Obsidian’s own table parser honor it correctly — confirmed directly in the Obsidian app, not assumed from the spec (an earlier version of this doc claimed otherwise from a test that turned out to be broken for an unrelated reason — a missing blank line before the table, not the escape itself). Outside a table cell (prose, the fold), | isn’t special and never needs escaping.
Inline concept mentions — in Tags and in prose — both use plain wikilinks. [[slug|Display text]] in prose, [[slug\|Display text]] in the Tags cell. The crawl-links plugin resolves the short slug (markdownLinkResolution: shortest) to the real Concepts page without needing the full path — except for a Concept folder’s own index page (e.g. tagging something broadly “AI/ML” via concepts/ml-ai/index.md): the shortest-slug matcher only matches a file’s own basename, and an index file’s basename is index, not the folder name, so a bare [[ml-ai]] silently resolves to a nonexistent /ml-ai instead of erroring — confirmed by testing it directly, not assumed. Use the full path with an explicit /index for those: [[concepts/ml-ai/index\|AI/ML]].
Prose mentions and Tags-cell mentions render differently even though both are wikilinks: body prose (p a.internal:not(.tag-link) et al.) is flattened to plain colored text, no background, no underline — the LaTeX-hyperref convention, so a mid-sentence mention doesn’t outweigh the words around it. The Tags row is the one deliberate exception: custom.scss re-applies the pill look (background, padding, radius) there specifically, targeted by the same “exactly two body rows” heuristic above rather than by any class a wikilink can’t carry. Same markdown syntax, different visual weight, chosen by where it sits.
On screens up to 600px wide, the metadata table is constrained to the reading width. Each tag pill remains a single line and truncates its own label with an ellipsis when necessary; separate pills may wrap between pills. This avoids both mid-label wrapping and horizontal scrolling of the whole metadata table.
Deeper notes (Background/Summary/Discussion/References). An Obsidian callout with a custom type, not the built-in note: > [!emergence-brief]- More details. Two things this is deliberately not:
- Not a native
<details>/<summary>element. It looks identical on the built site, but Obsidian’s renderer doesn’t reopen markdown parsing after a blank line inside a raw HTML block the way this pipeline’s remark/micromark does — bold, lists, and$$...$$math all came through as literal unrendered source once tried this way (confirmed by opening the file in Obsidian directly, not assumed). - Not
[!note]. Obsidian doesn’t recognizeemergence-briefand falls back to its generic default callout look (still fully functional — foldable, correct markdown processing) — confirmed the same way. Using an unrecognized, project-specific type name means the flat styling inquartz/styles/custom.scss(.callout[data-callout="emergence-brief"]) only ever touches this exact use, not any future genuine[!note]the site might add elsewhere.
A callout is a blockquote at heart — genuine markdown, not a raw-HTML island — so everything inside it (bold, lists, links, inline HTML <a> tags, KaTeX) renders correctly in both places without special-casing.
A display equation may appear as a block within a list item, separated from its surrounding explanation by blank blockquote lines and indented as continuation content. Keeping it inside the list item preserves the visual separation while preventing KaTeX’s block margin from becoming a direct callout-grid row that leaves dead space when the fold is closed.
An enumerated set of items — conditions a construction must satisfy, tasks a method was evaluated on, mechanisms under comparison — is a real markdown ordered list, not (1) ... (2) ... (3) ... run into one sentence: a nested 1. 2. 3. list under the introducing bullet, indented two spaces to stay inside the same list item, reads far better than the inline form once there are more than two items and renders correctly inside the fold.
Citations. Bracket markers [N], numbered by first appearance inside each item, not by the order references happen to get added to the list. Because numbering restarts while HTML IDs must remain unique across the whole page, every citation/ref pair carries a stable item key: <subject-code>-<item-label>, such as q-p1. The adopted subject codes, matching the seven reserved field values in DESIGN.md (resolved 2026-08-10): q for Quantum Information & Computing, qsm for Quantum Sensing & Metrology (first used 2026-08-13), mb for Many-body/Materials, ml for ML/AI, nmo for Numerical Methods & Optimization, dee for Electronic Devices & Engineering (renamed from the earlier dp/Device Physics code — no published item used dp yet, so this is a clean rename, not a migration), and mt for Mathematics. All seven codes are reserved now even though two (dee, mt) have no items yet, so a future item doesn’t need a new code added first — only a genuinely new eighth field would. The first occurrence of a citation therefore carries id="q-p1-cite-N" (only the first — a repeated citation’s later mentions link to the reference but don’t repeat the id); every occurrence links href="#q-p1-ref-N". The References list itself is a plain markdown list — no wrapping <div> to hang a class on, for the same raw-HTML-breaks-in-Obsidian reason as the fold above — with each entry hand-writing its own [N] as <a id="q-p1-ref-N" href="#q-p1-cite-N">[N]</a> instead of relying on <ol> auto-numbering, so the number is itself a link back to where it’s cited. The item key must be unique within the page and must not depend on title wording that may later change. Its default list bullet is hidden by .callout-content > ul:last-of-type in custom.scss, which picks it out by position (always the callout’s last list) instead of needing a class.
This is the one piece of markup left as hand-written raw HTML rather than a lighter markdown form, on purpose: the site already has a real, enabled pandoc-style citation plugin (@quartz-community/citations, [@citekey] syntax against bibliography.bib/emergence-numeric.csl), but it numbers and lists citations once per page, in appearance order across the whole document. This doc’s References belong to one item’s own fold, restarting at [1] — a second item citing something would need its own [1], not a continuation of the first item’s numbers. A page-global tool doesn’t fit a per-item design, so this stays hand-written. It’s also the least repeated piece of boilerplate here: unlike the metadata table and Tags line, which appear on every item, a References list only appears on an item that has deeper notes at all.
Jump behavior. Clicking a citation or reference number needs to both open a collapsed fold and land focus on the target — plain scrollIntoView isn’t enough, and this site’s SPA router (quartz/components/scripts/spa.inline.ts) intercepts same-page hash links before the browser’s own fragment-navigation ever runs. revealAndScroll in that file walks up from the target, opens any collapsed .callout or <details> ancestor it finds, then scrolls and focuses. It runs both from the router’s click handler and, separately, once on cold page load — a direct link straight to #ref-3 (bookmarked or shared) arrives without ever going through a click.
Worked example. About Briefings renders “Attention Is All You Need” as an actual Briefing item example rather than merely describing one: paper heading, metadata table, and a More details fold containing Background, Summary, Discussion, and References. It follows the content level established by the original Kim/Shi/Kato/Albert chiral-central-charge Briefing: enough background to state the paper’s question, the central mechanism and equation, the reported results, a discussion of the contribution and limits, and item-local primary-source citations. Collection-role guidance stays outside the example instead of masquerading as discussion of the paper.
An A (Announcement) item uses the same heading and metadata-table shape but may omit the fold when the short public account is already complete. Do not treat an unverified candidate as a worked example merely because its markup renders.
See Site maintenance for the rest of the maintenance self-check.