Seeds are the articles the researcher already knows belong in the review. They do three jobs in this workflow. Search terms are drawn from them, every candidate the searches find is scored by its similarity to them, and the references they cite become a second source of candidates. Whatever the seeds leave out, the rest of the pipeline is likely to leave out too, so this is the step that most deserves the researcher's time.
Section 1What a seed is for
A good seed set is not a reading list of the most famous works. It is a sample of the kinds of work the review should find, in the languages, disciplines and periods the review should cover. For a comparative review that means seeds for each case as well as seeds on the general question, since the similarity scores in Step 3 can only reward what the seeds contain. Seeds need not have a PDF. A record with a good abstract still contributes to the scores, and a record with only a title still keeps that work out of the search results, where it would otherwise be counted as new.
Section 2The seed collections
In the worked example the seeds sit in four subcollections of 0 Seeds, one for each search fixed in Step 0. The agent read all four through the Zotero web API, listed every item with its metadata and attachments, and copied the PDFs it found into the workspace.
| Seed collection | Items |
|---|---|
| Post-Socialist Transformation · general | 53 |
| Post-Socialist Transformation · North Korea | 107 |
| Post-Socialist Transformation · China | 53 |
| Post-Socialist Transformation · Vietnam | 51 |
| Total | 264 |
| with a PDF | 119 |
| without a PDF (metadata only) | 145 |
00_seed-intake). The check for seeds among the search results in Step 3 used 251 distinct seed identities.Section 3Reference documents
The researcher's own writing on the topic is the clearest statement of what the review is for. The worked example added two such reference documents to the seeds, a conference draft from September 2026 and a grant proposal to Taiwan's National Science and Technology Council. They are not part of the corpus and are never filed in Zotero. They enter only the comparison corpus, where they pull the scores towards the questions the researcher is asking.
Too little weight and the reference documents disappear among a hundred seeds. Too much and the review finds only works that resemble the researcher's draft. The worked example gave the seeds half of the total weight and each reference document a quarter. Each document therefore counts for as much as half of the seed set.
Section 4The comparison corpus
The agent turned the seeds and reference documents into a single table, the comparison corpus, with one row per document and the text the scoring would read. Seeds with a PDF contributed their full text and seeds without one contributed their abstract where Zotero had it. Three PDFs that were image-only scans, with no text layer, were used by title alone, and duplicate PDFs of the same work were set aside. The comparison corpus was prepared again on 12 October 2026, and the rerun replaced the first version.
| Component | Records | Weight each | Share of weight |
|---|---|---|---|
| Seeds with a full text | 112 | 1 | 50% |
| Seeds with an abstract only | 7 | 1 | |
| Conference draft (September 2026) | 1 | 59.5 | 25% |
| Grant proposal | 1 | 59.5 | 25% |
| Total | 121 | 100% |
06_prepare-comparison-corpus, which superseded 01_prepare-comparison-corpus. Each reference document weighs 59.5, half the 119 seed records, so the two together equal the seeds.Every score in Steps 3 and 4 measures similarity to this table. If one case has many more seeds than another, as North Korea does here, its literature will score higher. The worked example answered this by searching each country separately and selecting by several routes in Step 5, but the first remedy is a balanced seed set.
- seed_intake_manifest.csvOne row per seed item, with its collection, identifiers and whether a PDF was found.
- comparison corpusOne row per seed or reference document, with its text and weight, read by every scoring step.
Terminology
- Seed
- An article the researcher places in the review at the start, used to derive search terms, to score candidates and as a source of citations.
- Reference document
- A text by the researcher, such as a draft or a proposal, that states the aims of the review. It shapes the scores but is not part of the corpus.
- Comparison corpus
- The weighted table of seed and reference texts that every candidate is compared with.
- Seed identity
- A distinct work among the seeds, after duplicate items are matched. Search results that match a seed identity are not counted as new.