---
title: Perception Evaluation
description: Test what an agent notices — attention focus and detail orientation.
---

**Perception evaluation** tests whether an agent correctly identifies relevant information, focuses attention on the right aspects of input, and detects important details.

## What it tests

- Does the agent focus on the right aspects of the input?
- Can it distinguish signal from noise?
- Does it notice important details others might miss?
- Is the level of detail orientation appropriate for the task?

## Scorer: `perception_focus`

The `perception_focus` scorer evaluates whether the agent identified key elements in the input:

```go
testcase.ScorerConfig{
    Name: "perception_focus",
    Config: map[string]any{
        "expected_focus":  []string{"security", "performance"},
        "signal_keywords": []string{"injection", "XSS", "latency", "bottleneck"},
        "noise_keywords":  []string{"formatting", "typos", "style"},
    },
}
```

## Scenario: `perception_test`

The `perception_test` scenario generator creates test cases with a mixture of signal and noise to test attention:

```go
Case{
    ScenarioType: testcase.ScenarioPerceptionTest,
    Input:        "Review this code. It has some formatting issues, a typo in a comment, and a potential SQL injection in the user input handler.",
    Context: map[string]any{
        "signal": []string{"SQL injection"},
        "noise":  []string{"formatting", "typo"},
        "expected_focus": "security",
    },
}
```

## Perception parameters

| Parameter | Range | Description |
|-----------|-------|-------------|
| `AttentionFilters` | keyword/pattern list | What the agent should watch for |
| `ContextWindow` | 0.0 - 1.0 | How much surrounding context to consider |
| `DetailOrientation` | 0.0 - 1.0 | High-level overview vs fine-grained analysis |

## Dimension score

```go
result.DimensionScores["perception"] // 0.0 to 1.0
```

## Use cases

- Verify a security-focused agent prioritizes vulnerabilities over style issues
- Test that a data analyst agent notices anomalies in datasets
- Ensure a code reviewer catches subtle logic errors amid formatting noise