Authority before relevance
Retrieve only records already approved for the user’s context. Ranking can never promote a draft into authority.
AI assistant architecture
Answers must remain attached to authority.
AVERELI’s future assistant will retrieve approved knowledge, preserve programme ownership, cite its evidence, and refuse when the platform cannot support a trustworthy answer.
Governed source programmes6
Public indexed records4
Human escalationNOT_CONNECTED
Production claimsNone
Citations are mandatory
Every factual answer must resolve to source identity, owner, version, effective date, and a reachable canonical page.
Permissions travel with content
Public, authenticated, and restricted indexes remain isolated. Identity never implies product entitlement.
Instructions are untrusted data
Retrieved text, uploads, and user prompts cannot override system policy, access rules, or tool boundaries.
Refuse and escalate
When evidence is absent, conflicting, sensitive, or low-confidence, the assistant must say so and offer an accountable route.
Evaluate continuously
Grounding, citation accuracy, access control, privacy, injection resistance, refusal, and accessibility require recorded evidence.
Answer pipeline
Trust is checked at every boundary.
Retrieval is filtered before generation. Support is checked after generation. A fluent response without attributable evidence is a failure, not a partial success.
- 01Classify intent
- 02Authorize scope
- 03Retrieve governed records
- 04Generate with citations
- 05Verify support
- 06Answer or refuse
Knowledge boundary
AI is an interface, never a source of truth.
P100’s Engineering Knowledge Framework remains authoritative for engineering knowledge identity, classification, metadata, lifecycle, quality, and traceability. P003 consumes that contract and must preserve it through retrieval and citation.
Governed search →Security architecture →Contact and escalation →