Context Selection Checklist

Screen every candidate item

Include material because it supports the next decision, not because it is available. Prefer authoritative and current sources, preserve provenance, and include credible contradictory evidence that could change the result. Remove duplicates and superseded versions.

Classify sensitivity before sending anything to a model. Exclude secrets and data that is unnecessary, unauthorized, outside the approved retention boundary, or unsafe for the selected product. Mark retrieved pages, user text, documents, and tool output as untrusted data unless policy says otherwise.

Record a context manifest

Use one row per item. Summarize only after recording the source and omitted detail, so a reviewer can reconstruct what shaped the answer.

Decision: [NEXT DECISION]
Item: [SAFE LABEL — NO SECRET VALUE]
Source and owner: [SYSTEM OR DOCUMENT OWNER]
Authority: [PRIMARY | SECONDARY | UNVERIFIED]
Verified on: [YYYY-MM-DD]
Sensitivity: [PUBLIC | INTERNAL | RESTRICTED]
Trust: [TRUSTED INSTRUCTION | TRUSTED DATA | UNTRUSTED DATA]
Relevance: [WHY NEEDED NOW]
Contradiction value: [WHAT ASSUMPTION IT MAY DISPROVE]
Action: [INCLUDE | SUMMARIZE | RETRIEVE LATER | REJECT]
Omitted detail: [WHAT WAS LEFT OUT AND WHY]

Expected result and verification

The expected result is a bounded working set containing the instructions, evidence, examples, and current state needed for one decision, with each item traceable to its source and trust class.

Verify that every included item affects an objective or check, every volatile fact has a review date, contradictory evidence is visible, and restricted data has an approved purpose and destination. Ask whether removing each item changes the decision; remove it if not.