Context:
Consider the following {granularity}(s), which occur before the {granularity} to be analyzed:
{context_window}

Task:
Review the extracted themes for the following {granularity}:
{item}

Previously extracted themes:
{prior_relationships}

Instructions:
1. Review each extracted theme and its evidence against the text.
2. Keep themes that are clearly supported by the text.
3. Revise wording where needed so that each theme and evidence phrase is clear, concise, and faithful to the text.
4. If two themes are duplicates or substantially overlap, merge them into one.
5. If one theme combines distinct ideas, split it into separate themes.
6. Drop themes that are unsupported, incorrect, or redundant.
7. Add missing themes only if they are clearly and directly supported by the text. Do not infer or speculate.
8. Write each theme as a noun phrase or short gerund phrase (3–10 words) that captures a pattern of meaning. Avoid vague labels.
9. Write each evidence phrase as a short, self-contained phrase (3–10 words) paraphrasing the relevant passage.
10. Preserve important qualifiers such as negation, degree, and temporal scope.
11. Do not introduce information not supported by the text.
12. Focus on themes expressed in the target {granularity}. Use the context only to clarify meaning, not to introduce new themes.
13. The number of themes should generally be similar to the original list unless there is a clear reason to change it.
14. Return the corrected full set using this JSON schema:
{json_schema}
Return only valid JSON and no additional text.
