surveylisted

# Survey ## Overview Produce a literature survey grounded in papers actually retrieved, never in model recall. The skill runs a four-stage pipeline — search, screen, extract, synthesize — and emits a full report: summary, audit trail, thematic synthesis, comparison table, per-paper notes, and gaps. Every citation carries a verifiable identifier (arXiv id or DOI) returned by a source. The methodology follows Kitchenham's systematic-review phases for *conduct* and PRISMA's flow counts for the *audit trail*. The non-negotiable rule is citation discipline: LLMs fabricate references at high rates (GPT-4-class models fabricate roughly 18-20% of citations and get errors in another ~25-45% of the real ones), so this skill cites only what the fetch step returned. ## Modes Select with the first argument; auto-detect if omitted. - **discover** — given a topic or question, search the sources, screen for relevance, then synthesize. Use when the input is a subject ("survey long-context attention methods"). - **curated** — given a specific set of papers (arXiv ids, DOIs, or titles), fetch their metadata, then extract and synthesize from just those. Use when the input is a list the requester already trusts. Both modes share the extract and synthesize stages; they differ only in how the paper set is assembled. ## The fetch helper `scripts/paper_search.py` (stdlib only) is the grounding layer. Always fetch through it — do not write citations from memory. ```bash # discover py