Collecting Sources

Phase 1 of two-phase research: discover, evaluate, and store sources with metadata. Synthesis (Phase 2) and fact-checking (Phase 3) are separate phases — this skill ends at handoff.

Everything lands in the task workspace:

work/<task-id>/
├── source_index.md      # human-readable catalogue
└── sources/
    ├── papers/          # academic papers (.txt from CORE/Unpaywall, .pdf from arXiv)
    ├── web/             # web pages (Markdown)
    ├── blogs/           # blog posts (Markdown)
    └── docs/            # documentation (Markdown/HTML)

Every source file has a .meta.yaml companion conforming to schema v1.0 — fields, missing-data conventions, and a complete example in references/source-metadata-schema.md.

1 — Setup

• mkdir -p work/<task-id>/sources/{papers,web,blogs,docs}
• Source IDs are sequential and zero-padded: src_001, src_002, …
• If source_index.md exists from a previous collection, load it for duplicate detection.

2 — Discovery

Detect the domain from the research question:

• Academic: "paper", "study", "peer-reviewed", "journal"; biomedical (clinical, drug, patient), CS/physics (algorithm, neural network, quantum), social sciences (policy, governance, economics); academic URLs given (arxiv.org, doi.org, pubmed.gov)
• General web: "tutorial", "guide", "how-to", documentation focus, no academic terminology

Academic research — discovery finds papers and extracts identifiers (DOI, arXiv ID, PubMed ID, exact title) for the retrieval chain; do not fetch full text during discovery. Run 3–5 targeted WebSearch queries mixing plain and site-specific forms:

• Biomedical: site:pubmed.ncbi.nlm.nih.gov [topic]
• CS/physics: site:arxiv.org [topic]
• Any domain: site:scholar.google.com [topic], [topic] research paper peer-reviewed

Collect candidate metadata: title, authors, publication date, venue, abstract excerpt, identifiers.

General web research — run 3–5 targeted WebSearch queries with alternative phrasings; target authoritative sources (official docs, expert blogs, reputable sites); use site-specific searches when the site is known (site:docs.python.org async). Collect URLs, titles, authors, dates.

Target: 15–25 candidates from diverse source types, to be filtered to 8–15 on quality.

3 — Retrieval

Before fetching each candidate, check for duplicates: URL match against existing sources, DOI match against doi fields in existing .meta.yaml files. If already collected, skip and log.

Academic papers — fallback chain

Try in order; record which step succeeded as retrieval_method:

1. CORE API — always first; 46M full texts as plain text, all domains:

• WebFetch https://api.core.ac.uk/v3/search/works?q=doi:[DOI] — or q=title:"[exact title]" if no DOI
• If the JSON fullText field is non-empty, save it as .txt — done. Otherwise continue.
• Rate limit (free tier): 1 batch or 5 single requests per 10 seconds.

2. Unpaywall MCP — open-access PDF discovery and extraction:

• unpaywall_get_fulltext_links(doi=...) → best_oa_location.url_for_pdf; without a DOI, find one via unpaywall_search_titles(query="exact title")
• unpaywall_fetch_pdf_text(doi=..., truncate_chars=50000) → save as .txt (typical papers run 20,000–30,000 chars; surveys 90,000–100,000)
• Rate limit: 100,000 calls/day. If no OA version exists, continue.

3. Domain repository:

• arXiv (CS, physics, maths): prefer the arXiv MCP server if available; else query CORE by arXiv ID or title; last resort WebFetch https://export.arxiv.org/pdf/[arxiv_id].pdf — this stores the binary PDF only (WebFetch cannot extract PDF text). Max ~4 requests/second.
• PubMed Central (biomedical): WebFetch https://pmc.ncbi.nlm.nih.gov/articles/PMC[PMCID]/ — full-article HTML, store as Markdown or HTML.

4. WebFetch the landing page — occasionally yields full text as HTML; at minimum, abstract and bibliographic metadata.

5. Metadata-only (paywalled, last resort):

• First search for open-access alternatives: site:arxiv.org [title], site:biorxiv.org [title], [author] [title] pdf.
• If unavailable: record all metadata from the landing page; set file_path: "unavailable" and content_hash: "unavailable"; list the attempted strategies in provenance.notes; flag "Metadata only (paywalled)" in the index.

Social sciences caveat: OA coverage is ~33% vs ~66% for STEM, and SSRN (1.74M preprints) has no API. Expect 50–60% metadata-only rates and say so in the collection summary.

Web pages, blogs, documentation

WebFetch (converts HTML to Markdown). Preserve HTML only when formatting is critical.

Storage

Classify source_type (academic_paper, web_page, blog, documentation) and store:

Generate the content hash — shasum -a 256 <file> → content_hash: "sha256:HEX" — and compare against existing hashes to catch duplicate content fetched from different URLs.

4 — Metadata and RADAR

Write src_NNN.meta.yaml beside each source file, conforming to references/source-metadata-schema.md: core fields, source-type-specific fields, provenance (collected_by, collection_method, collected_at, task_id, notes), and retrieval_method for academic papers.

Assess every source on five RADAR dimensions, scored 1–5 with brief notes:

• Relevance — how directly it addresses the research question
• Authority — credibility of the author/publisher
• Date — currency relative to the field's pace
• Accuracy — verifiability and citation quality
• Rationale — purpose and potential bias

reliability_score = (Relevance + Authority + Date + Accuracy + Rationale) / 25.0

Full scoring rubric: skills/verifying-claims/references/source-evaluation-radar.md. Paywalled sources are assessed on abstract, venue, and authors.

5 — Quality filtering

• reliability < 0.5 — exclude (delete files and metadata); document exclusions and reasons
• 0.5–0.6 — flag for review; include only if the source count falls short
• ≥ 0.6 (or the requested threshold) — include

Minimum 8 high-quality sources per research question. If short: broaden queries, try alternative tools, lower the floor to 0.5 if necessary, and document the limitation. Retain everything above threshold — more coverage helps synthesis; prioritise the highest scores there.

6 — Source index

Write work/<task-id>/source_index.md:

# Source Index

**Task**: [task-id] · **Research question(s)**: […] · **Collection date**: [YYYY-MM-DD]

## Summary Statistics
- Total sources: [N] — [breakdown by type]; papers: [X] full-text / [Y] metadata-only
- Retrieval methods (papers): CORE [n] · Unpaywall [n] · arXiv [n] · PMC [n] · WebFetch [n] · metadata-only [n]
- Average reliability: [score]; distribution: excellent ≥0.8 [n] · good 0.7–0.79 [n] · acceptable 0.6–0.69 [n]

## Sources by Topic

### [Topic/theme]

**src_NNN** — [Title]
- **URL**: […] · **Type**: […] · **Reliability**: […] · **Retrieval**: […]
- **Summary**: [one or two lines]
- **Metadata**: `sources/<subdir>/src_NNN.meta.yaml`

## Collection Notes
- Tools used; quality threshold applied
- Limitations: [paywalled count, coverage gaps, domain caveats]

Group sources by themes extracted from abstracts/descriptions and the research question; note sources relevant to multiple topics.

7 — Handoff

Report to the orchestrator: total sources (full-text vs metadata-only), breakdown by type and retrieval method, average reliability, paths to source_index.md and the sources directory, exclusions and limitations (including domain caveats), and whether the collection is ready for synthesis or needs additional searches.

Constraints

• Working directory resets between Bash calls — use absolute paths.
• Bash only for hashing and file operations; no custom scripts.
• Out of scope: synthesis and analysis, fact-checking, research question formulation, MCP installation, paywall circumvention (no institutional credentials).

Reference APIs

• CORE API — core.ac.uk/services/api — 46M full texts as plain text in JSON
• Unpaywall — unpaywall.org/products/api — OA detection and PDF URLs
• arXiv — arxiv.org — 2.4M papers, direct PDF access
• PubMed Central — pmc.ncbi.nlm.nih.gov — biomedical full texts
• Semantic Scholar — semanticscholar.org — 214M papers, metadata only
• Google Scholar — scholar.google.com — academic search