Academic Research

Academic Research

Overview

Collects evidence from peer-reviewed papers and preprints. The defining risk here is exactly acceptance test 4 in this repo's test suite: a blog post misreading or overstating what a paper actually found. This skill's job includes resolving that conflict in the paper's favor when it arises — the root always outranks a derivative's misreading (engine/conflict-resolution.md's derivation check).

Layer rule: produces evidence, not results. Synthesis and confidence status are the orchestrator's job.

When to use

Direct use: "find studies on X", "what's the peer-reviewed evidence for Y". As a component: orchestrator dispatches this whenever a fact-type claim has a scientific/technical basis worth checking against the actual literature rather than someone's summary of it.

Don't use for: a secondary summary/interpretation as if it were the paper itself (fetch the paper directly, or flag when only the summary was accessible), general web content with no scholarly indexing (→ web-research).

Workflow

1. Search the actual literature index, not general web search — arXiv, PubMed, DOI resolution, or whatever academic search surface is available. General web search will miss paywalled or properly-indexed papers a dedicated search catches, and will surface blog summaries instead of the source.
2. Distinguish peer-reviewed from preprint from non-reviewed. A preprint (arXiv, bioRxiv, medRxiv) is real research not yet vetted — usable, but tier 2, not tier 1. Never present preprint findings with peer-reviewed confidence.
3. Read past the abstract for load-bearing claims. Abstracts compress and sometimes overstate; a claim central to the finding (a specific number, a causal claim, a stated limitation) should be checked against the actual body text before being recorded as the paper's claim — this is the single most common source of subtly wrong numbers in literature synthesis. Skimming the abstract is acceptable only for context/orientation claims, not for anything a fact finding will lean on.
4. If a search returns nothing: run references/query-relaxation.md, adapted to academic search syntax (drop field/date filters first, then reduce to core terms). Log full exhaustion as a gap — a thin-coverage area (e.g., non-arXiv CS venues, non-biomedical topics) is a real, reportable limitation, not a search failure to hide.
5. Check for retraction/correction status before recording — a retracted paper is excluded outright, not just tiered down; a correction changes what the record should quote.
6. Extract publication date — journal/preprint metadata is normally structured and reliable. Use published_at for the version you actually quoted, and modified_at when a revised version exists (a preprint later published with different numbers is a revision worth recording, not a footnote).
7. A paper is its own root: root_id: doi:<doi>. The preprint and the published version of the same work share it — quoting both is not corroboration. Independent corroboration means a different study, ideally by a different group.
8. Archive per references/citation-rules.md — for papers, a DOI is itself a durable identifier; still archive the specific page/PDF fetched, since DOI resolvers can change what they point to.
9. Assign tier per references/source-tiers.md's academic table (1 = peer-reviewed; 2 = preprint; 3 = thesis/workshop paper without full review; 5 = predatory-journal/retracted) and purpose (usually fact, sometimes implementation for methods papers with reproducible results).
10. Write the evidence record per engine/evidence-contract.md, verification_status: pending. Do not self-verify.

Output

One evidence record per claim, appended to the run's evidence.jsonl.

Common mistakes