The right model, agent or human for every request. With the reason.

# targets.yaml - one catalogue for every destination
targets:
  - id: llm-frontier
    kind: llm
    capabilities:
      domains: [math, coding, legal, finance]
      actions: [reasoning, planning, code_generation]
      min_complexity: 0.5
    constraints: { pii_allowed: false }
    cost: { usd_per_1k_tokens: 0.010 }
    latency_ms: 1800
    quality_prior: 0.92

  - id: coding-agent
    kind: agent
    capabilities:
      domains: [coding]
      actions: [code_generation, code_review, debugging]
    cost: { usd_per_1k_tokens: 0.004 }
    latency_ms: 4000

  - id: llm-onprem
    kind: llm
    data_boundary: on_prem
    regions: [eu, in]
    constraints: { pii_allowed: true }

# rules.yaml - preferences on top of hard policy
rules:
  - name: pii-stays-onprem
    when: { contains_pii: true }
    prefer: [llm-onprem, human-escalation]
    avoid: [llm-frontier, llm-mid, llm-small]
    weight: 1.0

  - name: hard-coding-to-agent
    when: { domains: [coding], min_complexity: 0.45 }
    prefer: [coding-agent]
    weight: 0.7

  - name: escalation-intent
    when: { actions: [escalate] }
    prefer: [human-escalation]
    pin: true

{
  "target": { "id": "coding-agent", "kind": "agent" },
  "confidence": 0.87,
  "elapsed_ms": 4.2,
  "signals": {
    "task_type": "coding", "complexity": 0.71,
    "contains_pii": false, "jailbreak_risk": 0.02
  },
  "policy_rejections": {
    "llm-small": "complexity 0.71 above max_complexity 0.45"
  },
  "ranked": [
    { "id": "coding-agent",  "utility": 0.91 },
    { "id": "llm-frontier",  "utility": 0.78 },
    { "id": "llm-mid",       "utility": 0.61 }
  ]
}