Literature review

A literature question has two halves: finding the papers a domain expert would point to, and turning them into something more useful than a reading list — a synthesis that says what's established, what's contested, what's new, and where the holes are. Both halves can fail quietly and look like competent output until someone checks.

Read the request for what it's actually asking

"What's the paper for X" wants one or two specific citations; "what's the evidence on X" wants a synthesis; "compare A and B" wants a comparison, not two adjacent summaries; "where are the gaps" wants the gaps, with the survey as supporting material. A two-word lay query wants you to choose the scope a domain expert would default to and say so up front — "I'll take this as asking about human RCT evidence; the animal literature is separate." Ask a clarifier only when the answer would genuinely change what you do.

Grounding: retrieve first, then write

For broad-survey, where-are-the-gaps, and compare-methods requests, the first move is a literature sweep — search_openalex / crossref_lookup from kernel.py, a PubMed query, web_search, or whichever domain connector is wired in (run search_skills({prefix:"mcp-"}) once to see which literature/data MCP servers are available, e.g. PubMed, Semantic Scholar, bioRxiv, ClinicalTrials.gov, and use the one that fits the field) — and the answer is built from what comes back. Your recall picks the framing; the retrieval picks the citations. A real survey usually carries on the order of fifteen or more distinct primary-paper DOIs, because each claim is anchored to the paper that established it; a handful of review citations is a reading list, not a synthesis. When the question is after a specific paper — "the original," "the seminal," a named trial or method — find the highly-cited primary publication that the follow-ups all cite, not a review or news piece about it.

That applies even when you know the answer cold. Resolving the DOI for a paper you're certain of — the Transformer paper, a textbook constant, a landmark trial — is a one-second tool call, and it's the difference between a citation and a claim about a citation. Verification is something that happens in your tool trace, not a sentence in your reply. A DOI you emit either resolves to a real paper that says what you claim, or it's a fabrication, and the difference is checkable in five seconds. When you have author/year/journal but not the DOI, look it up via CrossRef or OpenAlex rather than pattern-completing one; when even those details are hazy, that's a search query, not a citation. For recent developments, contested findings, or anything you "remember" from near or after your knowledge cutoff, retrieval isn't optional.

After the first sweep, take the two or three most relevant hits and walk one step in each direction on the citation graph: pull their reference lists (backward) and their cited-by lists (forward), then fold anything new and on-topic into the set before you start writing. The seminal paper a field builds on surfaces in the backward step; the recent work that extends or contests your top hits surfaces in the forward step, and neither reliably appears in a keyword sweep alone. expand_citations(doi) in kernel.py returns both directions from OpenAlex.

Retractions and the null result

Sensational papers are findable because they were sensational, and some were later retracted or failed to replicate. CrossRef's update-to field flags retractions; for any high-profile or surprising finding, a check takes seconds. The related trap is the question whose honest answer is "no such paper exists": when someone asks for "the paper showing X" and X fell apart or was never established, the right answer names the claim, says what happened to it, and points to what the actual evidence shows — not the closest-matching citation.