AI
Patterns worth testing

Staged model gateway: authenticate high-risk users, route by capability and policy, record the model and version, and enforce rate, cost and tool limits.
Memory quarantine: keep newly distilled skills in a non-executable tier and promote them only after provenance, safety and regression tests.
Multi-agent contracts: make role, inputs, outputs, authority, deadline and evidence requirements explicit.
RAG controls: source lineage, access-control lists, content sanitisation, freshness and retrieval confidence.
Observability: trace the task, model, prompt hash, sources, tool calls, approvals, cost, latency and recovery.




