Research with Evidence
Plan, execute, verify, and recover a bounded AI-assisted research workflow.
By the end
You will be able to
- Turn a research need into bounded questions and evidence requirements.
- Use AI to organize sources without treating generation as a source.
- Verify claim-to-source mappings and report disagreement or missing evidence.
- Recover from weak searches, inaccessible sources, and unsupported synthesis.
Prepare the question and evidence plan
State the decision the research supports, the questions that must be answered, the allowed time period, and the evidence strong enough for each claim. Separate facts, comparisons, estimates, and judgment.
Choose permitted data before searching. Do not upload confidential, personal, licensed, or organization-restricted material to a tool unless policy and authorization allow it.
Decision: [WHAT THIS RESEARCH SUPPORTS]
Questions: [BOUNDED QUESTIONS]
Scope: [TIME, REGION, PRODUCT, AUDIENCE]
Evidence required: [PRIMARY SOURCES, TESTS, DATA]
Allowed data: [CLASSIFICATION AND PERMISSION]
Stop conditions: [ENOUGH EVIDENCE OR ESCALATION]Execute with a source ledger
Search for primary sources, record canonical links and review dates, and note what each source can and cannot support. Use AI to extract candidate claims, organize themes, or identify gaps while keeping the original sources available.
Label generated suggestions as leads until verified. A model response, search snippet, or citation-shaped string is not evidence by itself.
Verify every material claim
Open each cited source and compare the exact claim with the supporting passage, table, test, or record. Check publication scope, version, date, and whether the source is authoritative for the claim.
For consequential conclusions, seek independent evidence or review. Report disagreements and uncertainty instead of averaging incompatible claims into false certainty.
Recover from evidence failure
If a source is unavailable, stale, circular, or unsupported, mark the affected claim and search for a stronger source. Narrow the question when the evidence cannot support the original scope.
When the evidence remains insufficient, deliver a gap report: what is known, what is disputed, what is missing, and the next safe action.
Protect data and preserve provenance
Minimize copied data, redact unnecessary identifiers, and preserve license and attribution requirements. Store source metadata separately from generated prose so a reviewer can reconstruct the reasoning.
Before sharing, check that citations resolve, sensitive information is absent, and the conclusion does not exceed the evidence.
Practice activity
Produce a one-page evidence-led research brief
- Choose a current, low-risk question and define its decision, scope, evidence requirements, allowed data, and stop conditions.
- Build a ledger with at least three primary sources, supported claims, limitations, and review dates.
- Draft a synthesis that maps each material claim to the ledger and labels uncertainty or disagreement.
- Remove one source, identify what no longer holds, and write the recovery action.
What to produce
- A research brief, source ledger, and one-page synthesis with claim-to-source mappings.
- A gap-and-recovery note showing the effect of the removed source.
Reflect before continuing
Which conclusion changed most when you tested whether its source actually supported its scope?
Evidence
Sources and verification
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileNIST · verified 2026-07-25
- Prompt engineeringOpenAI · verified 2026-07-25
- Prompting best practicesAnthropic · verified 2026-07-25
Knowledge check
Make it stick.
Choose the strongest answer for each question. Your attempts become part of your device-local transcript.