| name | research-corpus-review |
| description | Review a bounded paper corpus for papers relevant to a topic. Use when Codex needs to scan all or most papers from one or more named conferences, journals, proceedings sites, or other bounded sources, prove corpus coverage, rank papers by semantic relevance, inspect the source structure before choosing tooling, and shard large reviews across sub-agents without losing auditability. |
Research Corpus Review
Overview
Review a bounded paper corpus for a research topic without silently skipping papers. Start by writing a review plan, inspect the source structure, freeze a complete manifest, read every manifest entry at the appropriate depth, reconcile coverage, then report ranked results with an audit trail.
Keep automation and judgment separate. Use scripts to enumerate papers, normalize metadata, fetch artifacts, de-duplicate entries, and shard work. Read titles and abstracts yourself, and read deeper for high-signal or uncertain papers; do not outsource semantic relevance to keyword matching or brittle heuristics.
Workflow
- Write the review plan before touching the corpus.
- Capture the topic, sources, source types, year range, corpus boundary, expected deliverable, relevance rubric, and completion criteria.
- Record what counts as in-scope and what should be excluded.
- Read review-plan.md and use its template.
- Inspect the corpus structure before deciding how to collect it.
- Open the relevant source landing pages and identify where the corpus actually lives.
- Confirm how listing pages, pagination, issue pages, session pages, abstract pages, PDFs, and external links are structured.
- Decide whether direct browsing is sufficient or whether you need a site-specific script to build the manifest.
- Read site-inspection-and-manifest.md.
- Build and freeze the manifest before scoring relevance.
- Create a manifest artifact in the task workspace. Use CSV or JSON for large corpora; use a Markdown table only for small corpora.
- Give each paper one stable row and one stable identifier.
- Record the target corpus size and reconcile it against the collected manifest before reading for relevance.
- Do not start ranking until the manifest is frozen.
- Read every manifest entry and assign a structured result.
- Perform a first pass over every row: title, abstract, keywords, and any concise metadata needed to disambiguate the paper.
- Perform a second pass for high-signal or uncertain rows: introduction, conclusion, figures, and tables as needed.
- Perform a deeper read only for the top bucket or when the user asks for it.
- Read reading-and-ranking.md.
- Use sub-agents only after the manifest exists.
- Keep corpus definition, manifest generation, sharding, merge, and final reconciliation with the main agent.
- Assign sub-agents disjoint manifest shards with fixed output fields.
- Require one result per assigned manifest row. Missing fetches must be marked explicitly, not dropped.
- Read subagent-protocol.md.
- Reconcile coverage before claiming completion.
- Confirm that every manifest row is in a terminal state.
- Confirm that the terminal-state counts add back up to the manifest size.
- Call out duplicates, fetch failures, and unresolved rows explicitly.
- Read reporting-and-reconciliation.md.
- Report the ranked outcome and the coverage proof together.
- Return the corpus definition, manifest size, coverage summary, ranked list, deeper-read candidates, and any unresolved gaps.
- Do not return only the top papers; include enough audit detail to prove nothing relevant was silently skipped.
Operating Rules
- Freeze the corpus boundary early. Do not change the sources, year range, or inclusion rules mid-review without stating the change.
- Treat coverage as a first-class requirement. No coverage proof means the review is incomplete.
- Use the browser to understand the corpus structure; use scripts to mechanize repeated, low-judgment collection steps.
- When interactive browser exploration stabilizes into repeated collection steps, save the final automation code in the task workspace and record how to rerun it.
- Do not let scripts decide semantic relevance from token overlap alone.
- Escalate uncertain boundary cases into an explicit bucket instead of forcing false precision.
- Prefer a small number of durable artifacts over chat-only reasoning. The plan, manifest, and reconciliation summary are part of the deliverable.
Read These References When Needed
