Skip to content

Chore/site citation ci - #2

Merged
vinthony merged 4 commits into
mainfrom
chore/site-citation-ci
Aug 31, 2026
Merged

Chore/site citation ci#2
vinthony merged 4 commits into
mainfrom
chore/site-citation-ci

Conversation

@vinthony

Copy link
Copy Markdown
Member

Pull Request

变更内容

为什么应该收录

来源

检查清单

  • 链接指向可靠的原始来源
  • 已区分论文结论、官方声明和个人判断
  • 没有添加来源不明的模型或数据
  • 已检查 Markdown 和相对链接
  • 如涉及模型,已检查代码、权重和许可证状态

vinthony and others added 4 commits August 31, 2026 12:56
Nine references under resources/ and sources/ still pointed at assets/... as if
resolved from the repository root, so they rendered as broken images on GitHub.
Re-anchor them relative to the referring file.

Four diagrams under assets/imagegen-diagrams/001-004 were committed under the
raw image-generator output name (prompt text plus a timestamp, in CJK). Those
paths break on non-UTF-8 and case-insensitive filesystems and are unusable as
stable URLs; rename them to diagram.png, matching the other 89 folders.

Also track assets/zh/video-reasoning-roadmap.png, which docs/video-reasoning.md
already references but which had never been added, so that image was broken for
everyone except on the authoring machine.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ninety-odd chapters were readable only by clicking through GitHub's file tree,
with no full-text search and no stable URLs to cite.

MkDocs refuses to use the directory containing mkdocs.yml as docs_dir, but this
repo is read on GitHub as much as on a site, so every link is written relative
to the repository root. scripts/build_site.py hard-links the tree into
.mkdocs-build/docs before building, preserving that shape exactly, so one set of
relative links works in both places and no Markdown has to be rewritten.

Includes Chinese full-text search (jieba), light/dark themes, edit-on-GitHub
links, and MathJax wired through pymdownx.arithmatex so both the ```math fences
and inline $...$ render. MathJax is re-typeset via Material's document$
observable rather than DOMContentLoaded, which would leave equations unrendered
after an instant-navigation page change.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
bibliography/registry.json is machine-built from arXiv and Crossref, so those
132 entries are correct by construction, and lychee already catches dead URLs.
Neither covers the ~2800 hand-written per-chapter reference entries, where the
title, authors, venue or year can be wrong while the URL still returns 200.
That is the failure mode a reader would actually notice.

scripts/verify_citations.py checks arXiv links (anchor title against the real
arXiv title; claimed year against the v1 announcement) and CVF/ICLR/NeurIPS/ECVA
links (venue and year encoded in the URL path against the venue and year claimed
on the line). Claims are only attributed to a link on lines carrying exactly one
reference, since table rows list several papers from different venues side by
side.

A failed API batch is reported as unverified rather than as a missing record,
so a network outage cannot masquerade as hundreds of bad citations.

Two companion tools ship alongside: fix_asset_links.py normalises
generator-written filenames and re-anchors relative image paths, and
optimize_assets.py converts diagram PNGs to WebP (measured at ~17% of the
original bytes at equal or better fidelity than a 256-colour PNG). Neither runs
automatically.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The prose is Chinese, which is the point, but the repository has no
English-language surface at all, so nothing describes it to a reader arriving
from an English search or an awesome-list.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@vinthony
vinthony merged commit af7e0f5 into main Aug 31, 2026
0 of 2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant