pdf-cascade

Skill : pdf-cascade

Purpose

Wraps the paper-trail worker B's acquisition cascade. Given a single reference slug or a state filter, it advances the matching refs from candidate toward page1_validated through the FSM, with strict page 1 anti-homonymy validation.

Anchors all downloads in the local registry (pdf_path, pdf_sha256, acquisition_attempts[]) so the curator can audit everything.

When to invoke

Trigger this skill for any of:

• The user wants to fetch a PDF for a ref by slug
• The user wants to push the whole batch of candidate or uid_resolved refs forward
• The user explicitly calls /paper-trail:cascade, /paper-trail:reactivate-ocr, or /paper-trail:status
• sota-writer sub-task needs PDFs acquired for its proposed candidates

Do NOT invoke for semantic decisions (is this citation correct?) — that belongs to sota-auditor.

How it works

The skill delegates to the worker B Python CLI:

# Single ref by slug
python -m pipeline run --ref <slug>

# Batch by state filter
python -m pipeline run --state candidate --limit 50

# Dry-run (no mutation)
python -m pipeline run --state candidate --dry-run

# Reactivate refs waiting for OCR
python -m pipeline reactivate-ocr

The CLI invokes the 8-source cascade (or 10 sources if RESEARCH_ENABLE_SHADOW_LIBS=1 — see DISCLAIMER.md). Each acquired PDF must pass page 1 validation (author + title similarity ≥ 0.3 + zero off-domain keywords) before being accepted into the registry.

Cascade order (default, without shadow libs)

1. Crossref OA       (DOI-based, open-access metadata)
2. arXiv             (preprints CS/math/physics/q-bio/q-fin/etc.)
3. OpenAlex          (cross-domain academic graph)
4. Unpaywall         (OA discovery, fallback)
5. HAL               (Hyper Articles en Ligne, French academia)
6. CORE              (UK-based open repository aggregator)
7. archive.org       (digitized books and articles)
8. WebSearch queue   (manual fallback — adds the ref to a queue for
                     human-driven search via Claude Code interactive)

If shadow libs are activated, sources 8 and 9 are inserted before WebSearch:

8. scihub_optin       (Sci-Hub multi-mirror)
9. annas_archive_optin (Anna's Archive via scidb DOI + title search)
10. websearch

Anti-homonymy safety net

Every successful PDF download passes through _save_and_validate which:

1. Verifies PDF integrity (magic bytes, page count > 0)
2. Extracts page 1 text via pdftotext
3. Compares to expected metadata:
   • Author surname must appear
   • Title similarity ≥ 0.3 (keyword-based)
   • Zero off-domain keywords (e.g., arachnology terms for a CS paper)
4. Sets state: page1_validated if all 3 pass; pdf_acquired if PDF structure OK but text not extractable (likely scan, will trigger OCR via awaiting_rtfm_ocr); quarantines if validation fails

Quarantined PDFs go to _registry/_quarantine/<slug>_HOMONYM_*.pdf with the suffix indicating the failure mode.

Output

The CLI prints a session recap:

Récap session : planned=N done=N pending=N blocked=N skipped_terminal=N

And the doctor runs in the end (unless --no-doctor) to flag any invariant violation introduced.

For each ref processed, the acquisition_attempts[] field is appended in its registry file, providing a complete audit trail.

Examples

Acquire one specific ref

User: "télécharge le PDF de arnold_1982" Skill: invokes python -m pipeline run --ref arnold_1982 -v

Run cascade on all candidates

User: "lance la cascade sur les 30 prochaines candidates" Skill: invokes python -m pipeline run --state candidate --limit 30

Dry-run to see what would happen

User: "qu'est-ce qui se passerait si je lançais sur tous les uid_resolved ?" Skill: invokes python -m pipeline run --state uid_resolved --dry-run

Reactivate refs after RTFM OCR completed

User: "reprise OCR sur les awaiting_rtfm_ocr" Skill: invokes python -m pipeline reactivate-ocr

Failure modes

• cascade_exhausted : all sources failed → blocked_human:cascade_exhausted, the ref needs human-driven action (e.g., contact author, institutional access)
• title_mismatch : downloaded PDF doesn't match expected metadata → quarantined + blocked_human:title_mismatch
• breaker_open : a source had ≥ 5 consecutive failures, it's temporarily disabled for this session (Couche 2 circuit-breaker)
• worker_crash : exception in a source's helper → logged in journal, ref left in its previous state

Each is logged in acquisition_attempts[] with verdict and reason.

Recovery when the cascade exhausts (NEW v0.3.11)

When a ref ends in blocked_human:cascade_exhausted_needs_manual, the pipeline writes _registry/_hints/<slug>.md listing what was tried. Don't grab the PDF outside the pipeline and dump it manually — that breaks acquisition_attempts tracking, page 1 validation, and quarantine. Use one of three proper entry points :

Path A — You know an OA URL

If a quick web search surfaces a working URL (HAL page, university deposit, NIME proceedings, author homepage), inject it via the slash command :

/paper-trail:inject-url <slug> https://example.edu/path/to/paper

This sets oa_url: in the ref frontmatter, unblocks the ref, re-runs the cascade. The manual_oa_url source is at the head of the cascade and benefits from the landing→PDF resolver (so a deposit page URL works — no need for the direct PDF URL).

Path B — paper-search MCP download_with_fallback

For paywalled-but-on-RG / scidb references, the paper-search MCP handles ResearchGate, Sci-Hub fallback, and other anti-bot sites that our internal cascade can't reach :

# Inside Claude Code, not bash :
mcp__paper-search__download_with_fallback(
    title="<title>",
    doi="<doi>",
    save_dir="/tmp/paper-trail-recovered",
)

Then move the returned PDF to 10_SOURCES/<domain>/Sources/ and set the ref's pdf_path to its relative path. The next pipeline run --ref <slug> will pick it up via the local-PDF step and apply page 1 validation.

Path C — You already have the PDF locally

Just place it in 10_SOURCES/<domain>/Sources/ and edit the ref's frontmatter :

pdf_path: 11_Biblio_MIR/Sources/Author_Year_Title.pdf

Then python3 -m pipeline run --ref <slug> will validate page 1.

What NOT to do

• Don't bypass the registry by directly writing the file path into a SOTA — the doctor will flag it as an unanchored citation.
• Don't skip page 1 validation by manually editing state: page1_validated in the frontmatter — homonymies will leak in.
• Don't reach for _check_local_pdf from custom Python code — use the three entry points above.