The short answer
Gate retrieved sources by identity, authority, version, effective period, last verification, supersession status, property or entity scope, and conflict state. Define what the workflow must do when freshness cannot be proven—refresh, abstain, escalate, or use a safe deterministic fallback—before the model sees the content.
Operational checklist
Mark your progress, then save a working copy. Selections reset when you leave this page. A checked box is not an approval or evidence of completion.
☐
Source classes and authorities inventoried
☐
Version/effective rules defined
☐
Property and access scope enforced
☐
Supersession and conflict checks tested
☐
Untrusted content boundary applied
☐
Abstain/refresh/escalate behavior verified
☐
Decision provenance retained
0 of 7 marked
Key takeaways
- Recent retrieval is not the same as current source content.
- Freshness depends on the decision and effective period.
- Stale or conflicting sources require governed behavior, not model guesswork.
Write a freshness contract by source class
Inventory leases, policies, vendor documents, work records, payment states, owner instructions, building information, and external reference material. For each, name the authority, stable identity, version mechanism, effective date, expected update signal, allowed age where meaningful, scope, and unavailable-source response.
Some facts need event-driven confirmation rather than a time-to-live. A payment state checked five minutes ago may still be uncertain; an executed lease may remain authoritative for years until amended.
Evaluate retrieval before generation or tool use
Treat retrieved text as untrusted data and preserve provenance. OWASP notes retrieval-specific trust and prompt-injection risks; its guidance is a security reference, not a completeness guarantee.
| Check | Pass evidence | Failure behavior |
|---|---|---|
| Identity | Stable source and object ID | Reject ambiguous match |
| Authority | Approved system/owner for this fact | Use alternate authority or escalate |
| Version/effective time | Current version covers decision time | Refresh or abstain |
| Scope | Correct property, entity, user, and purpose | Exclude out-of-scope content |
| Supersession/conflict | No later source or conflict resolved | Present conflict to reviewer |
| Integrity/trust | Protected provenance and safe content handling | Quarantine or restrict use |
Retain enough evidence to reproduce the decision context
Record source IDs and versions, retrieval time, gate results, excluded sources, conflicts, model/workflow version, and action outcome without copying unnecessary sensitive content. Test expiration, replacement, access revocation, missing sources, and malicious instructions.
Monitor how often workflows abstain or use stale-source exceptions. A high refresh burden may indicate the source integration or ownership model needs repair rather than a longer arbitrary age limit.
Edge cases
- A cached policy is newest by timestamp but not approved: authority wins.
- Two source systems disagree during migration: present the governed conflict.
- A source has no version field: use approved hash/snapshot evidence or treat freshness as uncertain.
Sources and references
Follow each source to check the underlying claim. Access checks and professional review are different steps.
1. Primary source · OWASP Foundation
RAG Security Cheat SheetRetrieved content creates trust-boundary, provenance, and prompt-injection risks; retrieved material should be treated as untrusted input.
Source checked 2026-09-18
Automated source-access check: 2026-09-18.
2. Primary source · National Institute of Standards and Technology
AI Risk Management Framework CoreThe voluntary AI RMF addresses governance, measurement, monitoring, incident response, recovery, and change management. It does not prescribe a property-management workflow.
Source checked 2026-09-18
Automated source-access check: 2026-09-18.
Continue the workflow
Choose the source of truth for property operationsAI model and provider change acceptance gateReview AI workflow failure modes before launchRevision history
2026-09-18
Initial Phase 4 operational article with a distinct evidence artifact, failure states, source limits, and AI-assisted technical review.