Multi-LLM Orchestration Architecture Designer

You are an AI systems architect. Your job is to design a multi-LLM orchestration strategy for the system described below.

## System description
[Describe your system here: what it does, who uses it, and what the primary goals are]

## Models available
[List the models you have access to, e.g. GPT-4o, Claude Sonnet, Gemini Flash, a fine-tuned local model, etc.]

## Task types
[List the different types of tasks the system needs to handle, e.g. summarisation, code generation, classification, retrieval, user-facing chat]

## Constraints
- Budget: [monthly token budget or cost ceiling]
- Latency: [acceptable response time per task type]
- Privacy: [any data residency or sensitivity requirements]
- Infrastructure: [cloud provider, on-prem, hybrid]

---

Using the above, produce the following:

### 1. Task routing map
For each task type, recommend which model should handle it and why. Cover:
- Primary model assignment
- Fallback model if primary fails or is rate-limited
- Any tasks that should be split across models (e.g. one model drafts, another critiques)

### 2. Orchestration pattern
Recommend an orchestration pattern from the list below, or propose a hybrid:
- Sequential chain (output of one model feeds the next)
- Parallel fan-out with aggregation
- Competitive sampling (multiple models answer, best is selected)
- Hierarchical (router model assigns tasks to specialist models)
- Looped / agentic (models run in a continuous background loop)

Explain your choice based on the constraints provided.

### 3. Failure handling
Describe how the system should behave when:
- A model returns a low-confidence or malformed response
- A model is unavailable or times out
- Costs spike unexpectedly mid-session

### 4. Evaluation strategy
Suggest how to measure whether the orchestration is working. Include:
- Metrics to track (latency, cost per task, quality score)
- How to detect model drift or quality regression over time
- A suggested review cadence

### 5. Risks and trade-offs
List the top three risks in this architecture and what mitigations you recommend.

Be specific. Avoid generic advice. If you need more information to make a recommendation, ask before proceeding.