Systematic Research

Retrieve real sources, preserve provenance, and apply paper-extract to accessible material. Never generate catalog metadata or paper claims from memory.

flowchart TD
    A[Discover engines and read help] --> B[Search independent available engines]
    B --> C{Request succeeded?}
    C -->|No| D[Diagnose dependencies, network, rate limit or query]
    D --> E[Retry boundedly or use callable search/browser tool]
    E --> F[Preserve actual tool outputs and primary evidence]
    C -->|Yes| F
    F --> G[Merge/dedup without changing raw outputs]
    G --> H[Verify identity against source API or primary source]
    H --> I{Identity established?}
    I -->|No| J[Flag affected item; continue verified items]
    I -->|Yes| K[Download/save accessible source]
    K --> L[paper-extract, bibliography, R3 verification]

R0: Retrieval and identity

1. List scripts/search-*.py and search-engines/; read matching engines' help. Pair domain and general engines where available. Do not send a free-text topic to a conference-listing engine. Use separate output directories for each run.
2. Install optional dependencies only when needed. Read retrieval-and-failures for setup, USENIX syntax, output contracts, retries and callable-tool fallback.
3. Inspect exit status. Failure is not zero hits. Nonzero invocations write search-error.json; successful invocations include status: success and count. Do not merge stale results left by failed invocations.
4. Preserve raw script/tool outputs. Merge and dedup by URL/DOI without inventing entries or fields. Each field must trace to a retrieved response.
5. Verify title, available authors and abstract against source API/primary source. Compare up to the first three authors, allowing fewer-author papers. Preserve the response, method and discrepancies. API failures are not matches.
6. The bundled verifier supports arXiv only. Run it on a COPY, retaining raw input. DOI/OpenAlex/web items use saved source responses and verification records; never send unrelated identifiers to the arXiv API.
7. Download or save actual PDF/HTML content before extraction. A search snippet cannot support design or evaluation claims. Flag abstracts-only access.

E: Source extraction

Discover the sibling paper-extract skill and recursively list its templates and references. Read its SKILL.md; follow E0, E1, E1.5 and E2 as requested. Do not assume a template path. Record claims with actual sections/pages and distinguish source claims from proposed designs or hypotheses.

Each knowledge item contains actual source material, tldr.md, l0.md, l1.md when required, and bibentry.bib. Use bibtexparser; citation metadata must trace to retrieval/verification evidence. Aggregate with collect-bib.py as needed.

R3: Completion

Run verify-results.py on the knowledge subtree and retain its output. Check source existence, required extraction files, citation URL/DOI and author identity. Spot-check 2–3 decompositions against their sources for system names, claims and section evidence. Correct errors before declaring the library complete.