03. The site

Non-negotiable: it works offline

The first reader will open it on a plane. So:

Static HTML, built by a script, opened from file://. Double-clicking site/index.html must work with no server and no network.
No external anything. No CDN scripts, no web fonts, no remote images, no analytics. System fonts.
No fetch() of local files. Browsers block it on file://. Data a page needs is inlined into that page at build time (a <script type="application/json"> block is fine).
Relative links only. Every link between pages is relative, so the folder can be moved, zipped or served from any path.
Search works offline. A small client-side search over an index inlined into the search page.
External links are marked (↗) and open the live site; the local copy is always one click away beside them.

Build

python3 tools/build.py turns sources/, data/, editions/, stories/, history/, signals/ and loose-ends into site/.
Standard library only. No pip installs. Write a small markdown renderer (headings, paragraphs, lists, tables, code, links, emphasis); the sources use nothing more exotic.
Deterministic. The same inputs build the same bytes.
A validator (tools/validate.py) that fails the build on: a broken relative link, an external script or stylesheet, a fetch( in any page, a credential-shaped string (see 05-house-rules.md), or a page without a provenance block.
Also emit site/llms.txt and a .md twin of every page, the way every site in the network does.

Pages

Page	What it shows	Written by
Front page	Today's edition (the latest one), with the lead story, the counts of what changed per site, and links to the loose ends and signals.	Journalist
Editions	One page per day that had changes, newest first.	Journalist
Stories	Longer pieces from the editions.	Journalist
The index	Every page on every site: filter by site, type, date and concept. Each row links to the local copy and the live page.	Librarian
Reading room	Every source file rendered as a readable page, with its provenance block. The "new since 18 September" list first.	Librarian
Concepts	The recurring ideas, and every page that uses each.	Librarian
Vaults	Every published vault across the network, what it holds, its read key if the site publishes one.	Librarian
History	The week's moment, lessons learned, decisions and contradictions.	Historian
Signals	Cross-pollination: site A has X, site B should know.	Guest desks
Loose ends	What was said and not done.	All desks
The newsroom	A picture of the desks and the flow between them, with what each desk read and wrote on the latest day.	Build, from the data
About and method	What runs, what is design, who reviews, how sources are frozen. An "honest sentence" like newsroom.sgit.ai's.	Editor

The provenance block

Every page that states anything about a source carries, visibly: the site, the live URL, the date and time the source was fetched, its sha256 (first 12 characters are enough on screen), and which desk wrote the page. A reader must be able to walk from any claim to the frozen copy it came from.

Look

Borrow the look of sgit.ai: warm off-white background, near-black ink, one blue accent, a serif for headings, system sans for text, monospace eyebrows. Readable on a phone with a 16px gutter and no horizontal scroll. A dark mode is welcome and not required for the first version.

Data files

File	What it is
`sources/sites/manifest.json`	Every fetched file: URL, local path, bytes, sha256, fetched time. Written by `tools/fetch_sources.py`.
`data/changes/YYYY-MM-DD.json`	The Librarian's changes for the day: per file, site, type, title, one-line summary, concepts.
`data/index.json`	Every page in the network, with type, title, date if known, concepts.
`data/concepts.json`	Concept name, one-line definition, the pages that use it.
`editions/YYYY-MM-DD.md`	Front matter: `date`, `desk`, `sources` (list of URLs), `reviewed_by`, `reviewed_on`. Then the edition.
`stories/.md`, `history/.md`, `signals/*.md`	Same front matter shape.
`data/loose-ends.json`	Where said, when, what, waiting on, status.