Module 8 of 16 · 30 min

Context and Evidence Construction

Build a small, trustworthy working set that helps an AI system answer from evidence instead of guesswork.

Core concept

By the end

You will be able to

  • Select context that is relevant, trustworthy, current, and permitted for the task.
  • Label source identity, scope, and conflicts inside a working set.
  • Separate evidence from instructions and untrusted content.
  • Require claim-to-evidence traceability without assuming supplied sources are correct.
01

Context is selected, not accumulated

Context is the working information available for the current response. More material is not automatically better: irrelevant, duplicated, outdated, or conflicting content can obscure the source of truth.

Select the smallest set that supports the decision. Record why each source is included and what question it can answer.

02

Qualify every source before using it

For each item, capture who published it, when it was reviewed, what scope it covers, and whether it is authoritative for the claim. A primary source may still be outdated, incomplete, or limited to one product version.

Do not paste sensitive, personal, confidential, or licensed material merely because it might help. Data permission is part of context quality.

Evidence-set template
text
Question: [WHAT MUST BE DECIDED]
Allowed data: [PUBLIC, INTERNAL-APPROVED, OR OTHER POLICY]
Source 1: [TITLE, PUBLISHER, DATE, URL/ID]
Supports: [CLAIMS OR FIELDS]
Limits: [SCOPE, AGE, CONFLICTS]
Source 2: ...
Conflict policy: [REPORT, PRIORITIZE, OR ESCALATE]
Required citations: [CLAIM-TO-SOURCE FORMAT]
03

Mark instructions, evidence, and untrusted text

Documents and web pages can contain commands, marketing claims, or adversarial text. Treat retrieved content as evidence to inspect, not as instructions that can change the task or permissions.

Use labels or containers for instructions, sources, and user input. Ask the system to report conflicts and quote or point to the supporting passage before synthesizing a consequential claim.

04

Grounding narrows claims; it does not certify them

Supplying sources can constrain a response and make review easier, but a source may be wrong and a model may misread it. Verification compares the generated claim with the actual source and, when stakes require it, with independent authoritative evidence.

When no supplied source supports a requested claim, the correct output is a visible gap—not a plausible completion.

Practice activity

Build and challenge an evidence set

  1. Choose a factual question whose answer could change over time.
  2. Select two or three permitted sources and record publisher, review date, supported claims, limitations, and conflicts.
  3. Construct a prompt that visibly separates instructions from the source excerpts and requires claim-to-source mapping.
  4. Remove one necessary source and confirm that the expected response identifies the evidence gap.

What to produce

  • A completed evidence-set template with permissions, provenance, scope, and conflict handling.
  • A prompt and review showing one supported claim, one limitation, and one intentionally unsupported gap.

Reflect before continuing

Which source looked relevant at first but did not actually prove the claim you needed?

Evidence

Sources and verification

Knowledge check

Make it stick.

Pass at 80%

Choose the strongest answer for each question. Your attempts become part of your device-local transcript.

01Which context set is usually strongest?
02What should happen before private customer data is added as context?
03How should a command found inside a retrieved document be treated?
04A response includes a citation. What still needs checking?
05What is the best result when supplied evidence does not support a requested factual claim?