| name | literature-review |
| description | Structured process for a literature review, search strategy, source vetting, and synthesis. Use when asked to review existing work on a control-research topic (prior protocols, attacks, monitors, safety results). Not for incidental background reading during a run. |
Literature review
Models are weak at this by default: searches are shallow, sources go unverified (including hallucinated citations), and the output reads as a source-by-source dump instead of a synthesis. Counter each failure mode deliberately. Write the result into workspace/references/ and cite it from the report.
1. Search strategy
Do not run one query and stop. Cast wide, then iterate.
- Multiple query variants. Phrase the topic several ways: synonyms, the technical term and the lay term, narrow and broad framings. Different phrasings surface different work.
- Iterate on results. Broaden when you get too little; narrow or add qualifiers when off-topic. Pull the terminology a subfield uses from the first results and search again with it.
- Cover several venues. For academic work: arXiv, Google Scholar, Semantic Scholar (via
WebSearch/WebFetch). For AI-control grey literature: LessWrong and the EA Forum (via WebFetch) plus lab blogs (Anthropic, Redwood, DeepMind, OpenAI). Control research lives heavily in preprints and blog posts.
- Snowball. From each key paper follow both its references (earlier work) and its citations (later work) to find the cluster you missed.
- Fan out. For a broad topic, spawn several search subagents in parallel, each on a different angle, then merge.
2. Source vetting
Vet every source before you rely on it.
- Confirm it exists. Fetch the page; verify title, authors, venue. LLMs hallucinate plausible citations, never cite a source you have not retrieved.
- Assess credibility. Note the type (peer-reviewed, preprint, blog, forum comment) and weight accordingly, but do not dismiss grey literature: in AI control much of the relevant work is preprints and blog posts.
- Check recency. Note dates; flag findings that may be superseded in a fast-moving area.
- Primary vs secondary. Trace claims to the primary source, not someone else's summary.
- Watch for conflict. Flag where sources disagree or where later work overturns earlier results.
3. Synthesis
Organize by claim or theme, not by source. A list of per-paper summaries is not a literature review.
- Group findings around the questions or themes that matter, drawing on multiple sources per theme.
- State where sources agree, where they contradict, and where there are gaps (open questions, untested claims) that your experiment could fill.
- Calibrate confidence: distinguish well-replicated findings from single-study or contested claims, and say which is which.
- Cite everything so each claim is traceable to its source.