> ## Documentation Index
> Fetch the complete documentation index at: https://docs.routor.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Understanding Routing Decisions

> What each field in a routing decision means and how to read them.

# Understanding Routing Decisions

Every request Routor handles generates a routing decision. This page explains what each field means and how to read them.

***

## Where to See Decisions

Two places:

1. **Response headers** on every API call - instant, inline with your request
2. **Playground** - visual breakdown with color coding and confidence bar

> 📷 **\[Screenshot: Routing decision panel in the Playground showing tier, model, confidence, chain, and cost]**

***

## The Decision Fields

### Model

The actual model that handled the request, shown as `provider/model-id`.

```
moonshot/kimi-k2.6
zai/glm-5.2
deepseek/deepseek-v4-pro
```

This is what you see in the `model` field of the response body and in `X-Routor-Model`.

***

### Tier

Which of the 5 tiers the prompt was classified into. See [The 5 Tiers](../tiers.md) for the full breakdown.

```
NANO | SIMPLE | LIGHT | STANDARD | COMPLEX
```

***

### Confidence

How certain the classifier was, from 0 to 1.

| Score          | Meaning                              |
| -------------- | ------------------------------------ |
| 0.75 and above | Clear classification                 |
| 0.55 to 0.74   | Reasonable match                     |
| Below 0.55     | Borderline, sat near a tier boundary |

***

### Method

How the tier was determined.

| Value   | Meaning                                                 |
| ------- | ------------------------------------------------------- |
| `rules` | Rule-based classifier, standard path                    |
| `llm`   | LLM-based fallback classifier (reserved for future use) |

***

### Fallback Chain

The full list of models Routor built for this request, in order of preference. The first one was tried first. If it failed, the next one was used.

```
1. zai/glm-5.2                  tried first
2. moonshot/kimi-k2.6           tried if #1 fails
3. google/gemini-3.5-flash      tried if #2 fails
```

***

### Cost Estimate vs Baseline

| Field         | What it means                                          |
| ------------- | ------------------------------------------------------ |
| Cost estimate | What this request actually costs on the selected model |
| Baseline cost | What it would have cost on Claude Opus 4.8             |
| Savings       | The difference, shown as a percentage                  |

***

## Why a Decision Might Surprise You

If a prompt is routed to a higher tier than expected, these are the most common reasons:

* **Code keywords detected** - even mentioning function names or syntax can bump the tier
* **Multi-step patterns** - numbered lists or "first... then..." patterns score higher
* **Long context** - a large system prompt or long conversation history pushes toward COMPLEX

Use the [Debug Endpoint](../api/debug.md) to see exactly which dimensions fired and what score they contributed. That tells you precisely why the tier landed where it did.