Notes for AI assistants (Quarto blog repo)

Rendering + caching (Quarto freeze)

This repo enables Quarto freezing by default (_quarto.yml and posts/_metadata.yml set freeze: true), so HTML can show stale code output even after scripts change.
To force a post to re-execute when rendering, use one of:
- Add freeze: false to that post’s YAML.
- Render with quarto render <post>.qmd -M freeze:false --execute --no-cache.
- Delete/move the relevant _freeze/posts/<post>/ directory and re-render.

Pandoc/Quarto treats @something as a citation key.
When quoting Twitter handles or similar text, use @handle to display a literal @handle without triggering citations.

Files:

posts/2025-10-19-forecasts-of-AI-growth.qmd (includes an embedded “Validation Checks” block)
tools/forecasts_test.py (prints a Markdown checklist used by the post)
posts/ai.bib (bibliography, including quote + growth_annual_excess)
references/text/<citekey>.txt (local fulltext snapshots)

What tools/forecasts_test.py checks:

Citekeys in the QMD exist in posts/ai.bib.
Forecast table rows have consistent column counts.
QMD table quote and growth cells match posts/ai.bib.
Bib quotes match local fulltext in references/text/<citekey>.txt for forecast-table entries when the local fulltext is “usable”.
- Some sources block scraping (Cloudflare, paywalls, etc.); those local files may contain “verifying you are human”/permission errors and are treated as missing/unusable.

tools/fetch_bib_texts.py can populate references/text/ from text_url/url fields (skips PDFs).
Some sites will still block automated fetches; in that case you may need a manual copy or a different accessible text_url.

arXiv: prefer HTML via https://ar5iv.labs.arxiv.org/html/<id> (good for quote-matching); PDF is https://arxiv.org/pdf/<id>.pdf.
PDF-first sources (BIS/ECB/working papers, etc.): tools/fetch_bib_texts.py --pdf downloads PDFs and extracts text via pdftotext (Poppler).
Twitter/X: avoid citation parsing of @handle in Quarto by writing it as @handle in quoted text.
Paywalls/bot checks (Bloomberg/OECD/IMF/etc.): the “fulltext” you fetch may be a Cloudflare/Akamai placeholder (“verifying you are human”, permission errors). Treat these as unusable and find an alternative mirror/cached copy (author site, arXiv preprint, institution PDF, etc.).