Skip to main content

Writing Evidence-Based Decision Memos

A decision memo is an interface between analysis and accountable action. It should let a skeptical reader distinguish what the data establish, what remains a hypothesis, what must not be inferred, which guardrails can veto action, and who will obtain the next evidence.

Prerequisites​

Complete observational product analysis and confounding. Do not promote an association to a fact about impact when summarizing.

Learning goals​

  • Separate facts, anomalies, hypotheses, and causal limits.
  • Make a recommendation conditional on guardrails and evidence maturity.
  • Record rejected alternatives and a no-action option.
  • Assign owners, due dates, and falsifiable follow-up evidence.

Decision contract​

Lead with the exact decision and due date. Then use eight stable sections: verified facts; anomalies; hypotheses; causal limits; guardrails; recommendations; owners; and follow-up evidence. Link to versioned evidence rather than pasting an unauditable screenshot.

The recommendation should say what to do now, why that action is reversible, and what would change it. In this fixture, the safe recommendation is to hold automatic expansion until the delayed-data window closes and to predeclare a guarded experiment. “Roll out because adoption rose” is unsupported. “Do nothing” is still an option and needs an explicit consequence.

Report inconvenient evidence. The duplicate is real even though it is metric-neutral. The restatement increases the descriptive change, but also demonstrates that first-close values are provisional. The accessibility guardrail is not measured, so it cannot be marked as passing.

Worked example​

sections = [
"verified_facts", "anomalies", "hypotheses", "causal_limits", "guardrails",
"recommendations", "owners", "follow_up_evidence",
]
print(f"sections={sections}")
print("recommendation_owner=core-product decision_due=2026-10-06")

Expected output:

sections=['verified_facts', 'anomalies', 'hypotheses', 'causal_limits', 'guardrails', 'recommendations', 'owners', 'follow_up_evidence']
recommendation_owner=core-product decision_due=2026-10-06

Limitations and common mistakes​

  • Executive brevity is not permission to delete uncertainty or contrary evidence.
  • A traffic-light color without text, threshold, and owner is inaccessible and unactionable.
  • “More data” is not follow-up evidence until population, metric, design, owner, and due date are named.
  • A recommendation can be rational under uncertainty without being presented as certain.
  • Decision logs should preserve later outcomes without rewriting the original evidence.

Practice​

Write a one-page memo for the fixture. Include the no-action option, one condition that reverses your recommendation, one unmeasured guardrail, and the exact evidence the experiment owner must return.

Next steps​

Project: Turn Event Data into a Product Review builds the complete local evidence bundle and accessible memo surface.

Primary sources​

Checked 2026-10-19: HM Treasury, The Green Book 2026; W3C WAI, Tables Tutorial; UK Analysis Function, Communicating analysis.