sn-research-synthesis

Research Synthesis

Produce {report_dir}/synthesis.md. It is responsible for turning scattered sub-reports into ajudgment layer that can be written into the final draft, not a collection of sub-report summaries, and not the final reader-facing report.

Task boundaries

Do:

• Return to the original question and determine which findings truly answer what the user needs to judge.
• Distill 2-5 main lines of judgment.
• Judge the strength of evidence and applicable conditions for each main line.
• Summarize cross-dimensional consensus.
• Mark key conflicts and explain their sources.
• Clearly retain uncertainties, information gaps, and conditions that could overturn the conclusions.
• Explain what these judgments mean for the user's original question.
• Provide writing main lines and chapter allocation suggestions for the final draft.

Do not:

• Do not rewrite sub-reports.
• Do not summarize paragraph by paragraph by dimension.
• Do not directly generate report.md。
• Do not package gaps as conclusions.
• Do not add key judgments not supported by the sub-reports.

Input

• {report_dir}/request.md
• {report_dir}/plan.json
• All {report_dir}/sub_reports/*.md

What the comprehensive judgment must answer

At least clarify these 7 things:

1. Main conclusion: what the most important final judgment is.
2. Strength of evidence: how confident each judgment is.
3. Cross-dimensional consensus: which conclusions are jointly supported by multiple dimensions.
4. Key conflicts: which facts, definitions, time points, or positions contradict one another.
5. Explanation of conflicts: whether conflicts come from time differences, statistical definitions, sample differences, stakeholder positions, or undetermined facts.
6. Uncertainties and gaps: which questions still cannot be answered robustly.
7. Answer to the original question: what these findings mean for the user's initial question.

Execution process

1. Return to the original question: from request.md and plan.json.research_goal extract the judgment that ultimately needs to be answered.
2. Read all sub-reports: capture only the findings, limitations, and risks in each dimension that truly affect the conclusion.
3. Distill main lines of judgment: compress into 2-5 main lines that can support the final draft summary.
4. Judge strength of evidence: label each main line with 高 / 中 / 低 confidence, and explain why.
5. Summarize consensus and conflicts: distinguish "stable conclusions" from "contested conclusions."
6. Explain sources of conflict: do not just write the contradiction itself; explain why the contradiction exists.
7. Retain uncertainty: clearly state which conclusions hold only with conditions attached, and which still need further investigation.
8. Land on the user's question: clarify what these judgments mean for the user's decision-making, choices, understanding, or subsequent actions.
9. Lay out the wiring for the final draft: Explain which judgments go into sections such as the summary, key findings, and risks and uncertainties.

If a critical gap is found that prevents the main line from holding, first return to the corresponding dimension to supplement the research, then synthesize again.

Output

Write to {report_dir}/synthesis.md：

# Synthesis

## 原始问题

## 主线判断

## 证据强弱

## 跨维度共识

## 关键冲突与解释

## 不确定性与信息缺口

## 对原始问题的回答

## 对终稿的结构建议

Quality thresholds

• 主线判断 must be able to support the final draft summary, rather than being a vague summary.
• 证据强弱 must not only give labels; it must explain why it is strong or weak.
• 跨维度共识 only write conclusions that are truly supported by multiple dimensions.
• 关键冲突与解释 must explain the source of the conflict, and must not merely list two statements side by side.
• 不确定性与信息缺口 must not be omitted, nor beautified into "to be further observed".
• 对原始问题的回答 must directly respond to what the user initially wanted to judge.
• 对终稿的结构建议 only provide section allocation suggestions, and do not write the final draft body on behalf of others.

Common failures

• Copying each sub-report into a summary paragraph.
• Only conclusions, with no judgment of evidence strength.
• Avoiding conflicts when they are seen, without explaining the reasons.
• Writing it in advance as a finished draft for readers.
• Forgetting to return to the user's original question.