Research Source Trail Worksheet

Use this worksheet when AI helps with research, source triage, summaries, comparisons, or briefings. The goal is to keep the source trail visible from first question to final handoff.

Use it while you work. A source trail created at the end is often just a reconstruction.

Confidence Labels

Use the same labels throughout the worksheet.

• Confirmed: directly stated by an inspectable source you opened and checked.
• Likely: supported by a weaker source, a single source, or a reasonable pattern across sources.
• Unverified: no source has been checked yet, the source cannot be reached, or the claim was produced by AI without support.
• Conflicting: credible sources disagree, or the same source points in more than one direction.

Do not promote a claim because it sounds polished. Move it up only when the source trail supports the move.

Frame The Question

Write the research question before collecting sources.

• Question:
• Intended audience:
• Decision or deliverable this supports:
• What would count as a useful answer?
• What is out of scope?
• What would make the answer unsafe to use?

Working boundary:

This research is for [audience] and will be used for [decision or deliverable]. It should not claim [out-of-scope or unverified items].

Capture As You Go

Record findings while you work, not after the draft is finished.

Use source identifiers that someone else can inspect: URL, document title, page, note name, section, or source owner.

Source tier hints:

• Primary: closest available source for the claim.
• Secondary: interpretation, summary, commentary, or downstream reporting.
• AI output: useful lead or draft text, not evidence by itself.

Triage Each Source

Decide whether the source belongs in the research packet.

Notes on weak or partial sources:

• Notes:

If two sources agree only because they share the same origin, count them as one source and flag the dependency.

AI-Specific Traps

AI can make weak research look finished. Watch for these patterns before drafting or handing off:

• Fabricated citations: the title, URL, page, or section does not exist.
• Citation mismatch: the source exists but does not say what the draft claims.
• Stale currency: the output uses words like current, latest, or now without a checked date.
• Confidence laundering: fluent prose makes an unverified claim feel settled.
• Over-synthesis: the draft removes caveats or disagreement to create a cleaner story.
• Single-origin collapse: several sources repeat the same original claim without independent confirmation.

Example:

Draft claim: “The standard onboarding window is 14 days.” Label: needs source. Action: capture the source, rewrite as an open question, or remove the number.

Source-Said Versus AI-Added

Review the AI output and label claims.

Rules:

• Source said: the source directly supports the claim.
• AI inferred: the claim may be reasonable but is not directly stated.
• Needs source: do not ship the claim until a source is captured or the claim is removed.

When in doubt, downgrade the label. A useful inference can still be included as a recommendation instead of evidence.

Uncertainty Pass

Before handoff, group claims by confidence.

Confirmed:

• Claims:

Likely but not fully proven:

• Claims:

Unverified:

• Claims:

Conflicting:

• Claims:

Decision:

The final deliverable may use [confirmed claims]. It must label or remove [likely / unverified / conflicting claims].

Verification Sign-Off

• Reviewer:
• Date:
• Sources checked:
• Claims checked:
• Claims removed:
• Claims still uncertain:
• Expert or owner needed:

Final confidence label:

• confirmed enough for intended use
• partial, with gaps listed
• unverified draft only
• conflicting evidence, do not use as recommendation
• needs expert review

If any shipped claim still has the label needs source, the packet is not ready. Cite it, label it as unverified, or remove it before handoff.

Worked Mini-Example

Scenario: a learner is researching whether a team needs a short internal guide for reviewing AI-generated meeting summaries.

Decision:

Use the owner/date and missing-decision claims with labels. Remove the rework percentage unless a real source is added.

Handoff note:

I checked the training note and synthetic exercise examples. Owner/date guidance is confirmed. Missing-decision risk is likely but based on exercise notes, not measurement. I removed the unsupported rework percentage.

Full Practice Packet

Use this packet when you want a realistic source-trail exercise instead of a blank worksheet.

Scenario

Riley is a new operations analyst at Northstar Harbor, a fictional regional logistics company. Her manager wants to know whether the service desk should pilot AI-assisted ticket summaries next quarter.

An AI draft recommends an immediate rollout and says the team can expect “about six hours per analyst per week” in savings. Riley needs to prove which parts of that statement are sourced, which parts are inferred, and which parts must be removed.

Research Question

Should the service desk pilot AI-assisted ticket summaries next quarter, and what guardrails should be in place before the pilot starts?

Source Packet

Source A: Service Desk Weekly Review, 2026-06-19

• Team leads reviewed 24 escalated support tickets from the prior week.
• 9 tickets had handoff notes missing either customer impact or next owner.
• 4 tickets had conflicting status language between the ticket summary and latest comment.
• The review did not measure time spent writing summaries.
• Team lead Priya marked the quality problem as “frequent enough to slow handoffs.”

Source B: Pilot Tool Vendor Overview, retrieved 2026-06-24

• The vendor says the tool can summarize ticket comments and propose next-step text.
• The vendor recommends human review before summaries are sent to customers.
• The page includes a customer quote about faster handoffs, but no baseline, sample size, or measured savings.
• The page does not mention Northstar Harbor or its ticketing workflow.

Source C: Internal Security Review Note, 2026-06-22

• Security approved testing with synthetic or redacted ticket examples only.
• Real customer identifiers, contract details, access tokens, and production incident logs are not approved for use in the pilot tool.
• Security asked for a named reviewer on any summary that leaves the service desk.
• Legal review is still pending for customer-facing use.

Source D: AI Draft Summary

AI-assisted summaries are clearly ready for rollout next quarter. Recent service desk reviews show summaries are slowing handoffs, and the vendor’s customers report faster work. The team can expect to save about six hours per analyst per week. Security has approved the pilot as long as a human reviews the output, so the service desk should start with live tickets and measure the results.

Traceability Task

Fill the table before writing the recommendation.

Labels:

• source said
• AI inferred
• needs source

Actions:

• use with source
• rewrite as recommendation
• downgrade to likely
• remove
• assign owner

Expected Answer Shape

The completed answer should include:

• at least six traced claims
• at least one removed claim
• at least one rewritten recommendation
• a confidence label for each important claim
• a short final recommendation that separates internal pilot readiness from customer-facing rollout readiness
• named human owners for quality criteria, data boundary, and customer-facing review

Sample Answer

Final recommendation:

Recommend a small internal pilot using synthetic or redacted ticket examples. The pilot should measure whether AI-assisted summaries help reviewers catch missing customer impact, missing next owner, and status-conflict issues already found in the weekly review. Do not use live customer tickets or customer-facing summaries until security and legal clear those uses.

Removed or downgraded claims:

• Removed the six-hour savings claim because no source measures time savings.
• Removed the live-ticket recommendation because Source C explicitly limits testing data.
• Rewrote customer-facing rollout as a blocked future decision because legal review is pending.

Owners:

• Priya owns service-desk quality criteria.
• Security owns the data boundary.
• Legal owns customer-facing clearance.

Rubric

• Strong: Claims are traced at the claim level; the six-hour savings claim and live-ticket rollout are removed; the recommendation clearly separates internal pilot, production data use, and customer-facing use.
• Adequate: The major unsafe claims are removed, but some confidence labels or owner names are too broad.
• Needs work: The answer repeats the AI draft’s confident rollout language, treats vendor marketing as proof, misses the synthetic/redacted data limit, or leaves customer-facing use unresolved.

Common Mistakes

• Treating an AI draft as a source instead of a work product to inspect.
• Counting a vendor quote as independent measurement.
• Keeping a specific savings number because it makes the recommendation sound concrete.
• Saying “security approved the pilot” without preserving the data boundary.
• Forgetting that legal review is pending for customer-facing use.

Self-Check Questions

• Which claim in the AI draft has no support in Sources A, B, or C?
• What is the difference between “pilot internally” and “roll out to live customer tickets”?
• Which source is closest to the service desk problem?
• Which source is closest to the data-use boundary?
• Who must approve customer-facing use?

Completion Evidence

Attach these items to your training record or manager handoff:

• completed source trail table
• final recommendation paragraph
• removed or downgraded claim list
• owner list for quality, security/data boundary, and customer-facing clearance

Handoff Gate

Before sharing the research output, confirm:

• The research question is visible.
• Important claims have sources.
• AI-added claims are labeled or removed.
• Weak sources are dropped or flagged.
• Uncertainty is visible.
• A human reviewer is named.
• The next person knows what to trust and what to verify.
• No final claim still says needs source.

Handoff note:

I checked [sources] and confirmed [claims]. The following items remain [likely / unverified / conflicting]: [items]. The next owner should [action].

Reference Links

These public references are useful starting points for deeper study. The worksheet above is original LIW training guidance.

• OpenAI: Prompt engineering
• Anthropic: Prompt engineering overview
• People + AI Guidebook