> ## Documentation Index
> Fetch the complete documentation index at: https://fpde-80-mintlify-48090872.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Explain one sample

> Fit FPDEEngine and inspect a single target-versus-rival explanation

Use `FPDEEngine.explain_one` when you want one attribution vector and the metadata behind it.
This is the common path for local explanations.

## Prerequisites

* A training matrix `X_train`
* Training labels `y_train`
* One explanation vector `x`
* A fitted classifier with `predict_proba` and `classes_`

Apply the same preprocessing to `X_train` and `x`.

## Fit reusable FPDE state

```python theme={null}
from fpde import FPDEEngine

engine = FPDEEngine.fit(X_train, y_train, model=model)
```

FPDE builds one class-mean prototype per class.
The engine stores those prototypes, labels, anchors, and the baseline vector.

## Explain the sample

```python theme={null}
attributions, details = engine.explain_one(
    x,
    lambda_hyb=0.5,
    normalize="l1",
    anchor_strategy="mean",
)
```

`lambda_hyb` controls the Hyb-FPDE mixture.
A value of `1.0` uses the Diff-FPDE endpoint.
A value of `0.0` uses the Cos-FPDE endpoint.

## Read the details

```python theme={null}
print(details["target_label"])
print(details["rival_label"])
print(details["target_probability"])
print(details["evidence"])
print(details["exactness_residual"])
```

| Field                | Meaning                                                               |
| -------------------- | --------------------------------------------------------------------- |
| `target_label`       | Model-selected target class.                                          |
| `rival_label`        | Model-selected rival class.                                           |
| `target_probability` | Probability for the target class.                                     |
| `lambda_hyb`         | Hyb-FPDE mixture weight used for the explanation.                     |
| `evidence`           | Target-versus-rival contrast decomposed by FPDE.                      |
| `exactness_residual` | Numerical difference between summed attributions and direct evidence. |

<Note>
  Positive attribution values support the target class.
  Negative attribution values support the rival class.
</Note>

## Show the strongest features

```python theme={null}
import numpy as np

feature_names = np.asarray(data.feature_names)
order = np.argsort(np.abs(attributions))[::-1]

for index in order[:10]:
    print(feature_names[index], attributions[index])
```

Use the raw attribution values for evidence-scale reporting.
Use normalized values only when you need a display scale.

## Common next step

After you confirm one explanation, use [Explain batches](/explain-batches) to explain evaluation or test samples with the same fitted engine.