ARIA Shield

How it works

1. Edge authenticates SPA with HTTP-only cookies, not tokens.
2. BFF validates session and requests PDP decision per route.
3. On allow, BFF brokers per-service tokens and applies constraints.
4. Receipts are signed and chained; budgets settle to actuals.

Budget semantics (HTTP 402)

HTTP/1.1 402 Payment Required
Content-Type: application/json

{
  "error": "budget_exceeded",
  "call_id": "abc-123",
  "policy": "llm.invoke",
  "remaining_usd": 0.00,
  "retry_after_s": 86400
}

When a route exceeds its budget, Shield returns a deterministic 402 with call_id and remaining budget. On permit, the hold is settled to actuals and linked in the receipt.

Streaming caps

// Pseudocode
const caps = { tokensMax: 2048, durationMsMax: 20000 };
shield.stream(model, request, caps, onChunk);

Streaming stops early when token or duration limits are reached as dictated by PDP constraints.