Engineering Blog

Enhanced AI Tokens for Australia: Architecture and Runtime

September 23, 20258 min read
Policy • Tooling • Locale

Enhanced tokens enrich standard LLM inputs with runtime policy, tool hints, and Australian locale metadata. This lightweight layer improves safety, utility and relevance—without modifying underlying model checkpoints.

Design Goals

  • Inject governance at the token boundary: redaction, classification, routing
  • Expose tool affordances (schemas, cost hints) for planner‑style models
  • Attach locale context (spelling, regulation, industry codes) for Australian outputs

Runtime Flow

  1. Parse request and map tenant policy to token metadata
  2. Annotate spans with safety labels and tool‑eligible fields
  3. Select model and sampling policy; stream responses with audit tags

Key Takeaways

  • Enhanced tokens add policy, tool and locale context without model changes
  • They enable safer outputs and better task completion rates
  • Australian spelling, regulation and industry norms are handled by design