lit-reviewlisted

# Literature review, end to end A review is finite only when it serves a decision. Name the decision first, or a citation-graph walk expands forever. The six stages below are ordered, and the order is the method. Each writes a file the next one reads, so the whole review is a folder someone else can re-run. ``` 1 clusters how humans describe the problem -> clusters.csv 2 seeds papers you already trust -> seeds.csv 3 harvest the citation graph, both ways -> candidates.csv + named.csv + overlap.md 4 shortlist screened, with provenance -> shortlist.csv + funnel.md 5 extraction quote-grounded columns -> shortlist_extracted.csv 6 experiments findings that can be wrong -> experiments.md ``` Scripts are stdlib-only Python 3.9+ and live in `scripts/`. Run them from the skill directory or copy them next to the review folder. OpenAlex and arXiv are the only services used; both are free and need no key. Set `LIT_REVIEW_MAILTO=you@example.com` for the polite rate-limit pool. ## Before anything else: what decision does this serve? Write one sentence in `review/question.md`: > This review exists to decide **X**, between options **A / B / C**, under > constraints **...** Then the stopping rule follows from it, and "have I read enough" stops being a feeling. A review with no decision to serve has no stopping rule; a review with one stops at saturation, when a hop returns no mechanism not already logged. State