ever-works/orgs/synthesis-memo

Synthesis Memo

Turn findings into an answer someone can act on. The failure mode this prevents: the research dump — twenty pages of everything learned, the answer buried on page fourteen, every claim equally confident regardless of what supports it.

Steps

1. Reread the research plan before the findings. The memo answers the plan's question for the plan's decision; everything else learned goes to an appendix or dies.
2. Synthesize upward in three layers, written bottom-to-top but read top-to-bottom:

• Findings — rated, sourced observations (already produced by the study).
• Patterns — clusters of findings pointing the same direction; a pattern needs at least two independent findings.
• Claims — the assertions that answer the question, each resting on named patterns.

1. Write the opening paragraph last, place it first: the question, the answer, and the confidence level in three sentences. A reader who stops here leaves correctly informed — that is the test.
2. Carry evidence ratings into the prose. "We observed", "practitioners report", and "the vendor claims" are different sentences and the reader must never need the appendix to tell which they got. Never let synthesis silently upgrade a rating.
3. Write the disconfirming section as a first-class citizen: findings that cut against the answer, and — as important — what was not found that should exist if the answer were true. An empty disconfirming section means the search for it was skipped, and says so.
4. State the residual: open questions, the segments or sources missing, and what new evidence would change the answer.
5. Size to the decision: a reversible bet gets one page; only an expensive, irreversible call earns five. Cut by moving detail down the layers into the appendix, never by cutting hedges the evidence requires.

Quality bar

• Answer in the first paragraph, confidence stated in words that mean something ("would survive one contradicting source" beats "fairly confident").
• Every claim traces to patterns and findings; an untraceable claim gets deleted or demoted to labeled speculation.
• The disconfirming section exists and is specific.

Output format

Opening (question, answer, confidence) → claims with supporting patterns → disconfirming evidence and absences → residual and change-triggers → findings appendix with ratings and sources.