# Model Escalation Policy (Excluding GPT‑5.1)

**Summary:**
Use local Qwen3‑4B + RAG as the default “brain.” Escalate to cloud models only when the task clearly needs more capability. Different models are best for different kinds of work and budget levels.

---

## 1. Default: Local Qwen3‑4B + RAG

**Use for ~90–95% of everything:**

- Daily chat and reasoning.
- Question‑answering over your own notes via ChatDistill (Qdrant).
- Light–medium coding help, refactors, and debugging.
- Infra questions about:
  - llama‑server flags,
  - your hardware profile,
  - RAG design and workflows.

**Why:**

- Runs fast on the P330 + P1000.
- Has your **ground truth** in RAG (hardware, tools, cost notes).
- Free (local) and “good enough” for most decisions that are easy to reverse.

---

## 2. When to Escalate to Qwen3 Coder 30B A3B

**Role:**
Primary **coding specialist** when you need higher quality than 4B.

**Use when:**

- You’re working on **complex, error‑sensitive code**, e.g.:
  - multi‑file refactors,
  - non‑trivial algorithms,
  - architecture‑level changes.
- You want more reliable:
  - type correctness,
  - API usage,
  - edge‑case handling.

**Why this before others:**

- Tuned for code; better coding ability per dollar than general models.
- Cheap enough (~$0.15 / 1M tokens) to use regularly for serious coding work.
- Best “bang for buck” when the bottleneck is **code quality**, not general reasoning.

**Policy:**
> “If 4B feels shaky on a code change I actually care about → escalate to **Qwen Coder 30B**.”

---

## 3. When to Escalate to GPT‑4.1 nano

**Role:**
Cheap, stronger **general reasoner** and analyst.

**Use when:**

- You want a **second opinion** on:
  - architecture or design decisions,
  - non‑code system planning,
  - trade‑offs inside your stack (pipelines, batching, hardware configs).
- You need more brainpower than 4B but don’t want to pay 5.1‑level prices.
- You’re doing **high‑volume** but not ultra‑critical analysis / coding.

**Why:**

- Strong general‑purpose reasoning and breadth for its price.
- Good “cheap reviewer” of Qwen3‑4B’s plans, as long as you still verify key numbers.
- Better default choice than OSS‑20B when you care more about **overall quality** than being strictly free.

**Policy:**
> “If a plan from 4B touches multiple systems or feels borderline, or I want a cheap audit → run it past **GPT‑4.1 nano**.”

---

## 4. When to Use OSS‑20B (Free)

**Role:**
Free, reasonably capable general model for **bulk / low‑stakes** work and rough analysis.

**Use when:**

- You need to generate **lots of text** where quality can be lower:
  - brainstorming,
  - rough drafts,
  - expanding simple notes.
- You want to offload work that would otherwise eat a lot of tokens, but where:
  - correctness is not critical,
  - and you’re happy to clean up manually.
- You want a **zero‑cost second opinion** and are comfortable sanity‑checking results yourself.

**Why (vs GPT‑4.1 nano):**

- **OSS‑20B:** free, sometimes good at detail/spec nitpicks, but less reliable than nano overall on hard reasoning.
- **GPT‑4.1 nano:** generally stronger and more consistent, but costs a little per token.

Use whichever matches your priority:

- If **budget = 0** and stakes are low → **OSS‑20B**.
- If **quality matters more than $0.10/1M tokens** → **GPT‑4.1 nano**.

**Policy:**
> “If it’s large‑volume and low‑stakes → use **OSS‑20B**.
> If I want a more reliable cheap second opinion → use **GPT‑4.1 nano**.”

---

## 5. Relationship to GPT‑5.1 (If/When Used)

Even though this note is about the *cheap* models, keep the mental slot:

- **GPT‑5.1 chat** – for **rare, high‑stakes** tasks only:
  - critical business/strategy decisions,
  - anything with legal/financial risk,
  - deeply non‑obvious or safety‑critical code.

**Policy:**
> “Use 5.1 only when a mistake would be expensive or dangerous.”

---

## 6. Escalation Decision Tree (Summary)

1. **Start with local Qwen3‑4B + RAG** for everything.
2. If it’s **serious code** and 4B feels shaky → **Qwen Coder 30B**.
3. If it’s **system design / planning / infra trade‑offs** and you want a stronger brain:
   - Prefer **GPT‑4.1 nano** if you’re okay spending a bit, or
   - Use **OSS‑20B** if you want a free but weaker second opinion.
4. If it’s **cheap, bulk, low‑stakes text** → **OSS‑20B (free)** by default.
5. For **critical, irreversible, or high‑risk decisions** → only then consider **GPT‑5.1**.

Sticky note version:

> **4B + RAG by default.
> Coder 30B for real code.
> 4.1 nano for better general audits.
> OSS‑20B for free bulk and low‑stakes checks.
> 5.1 only for truly high‑stakes.**