# User guide

## Command center
Select classifier, split and stress family. Observe clean accuracy, percentage of actually changed text, prediction flip rate and modified-subset accuracy difference. Watch the paired-bootstrap warning and inherited-label limitation.

## Stress matrix and paired examples
The matrix compares all conditions. The paired explorer shows each original text and its exact altered version and prediction. Search by example ID or text and filter to prediction flips.

## Selective risk and reliability
Risk vs. coverage ranks samples from strongest to weakest model-specific **top-two score gap**. Drag the slider to see observed error among retained examples. Naive Bayes reliability bins summarize posterior confidence; do not assume calibration. Similarity baselines have no reported probability calibration.

## Language slices and gates
Inspect support sizes for EN/ES/PT, domains or labels. Edit research thresholds in Research gates and see what fails. Passing all checks is not a production approval.

## Sandbox and export
Write 12–4,000 characters, choose language and a stress family, then click **Compare locally**. Without a known gold label, this only compares two predictions. Download Markdown, CSV, or JSON reports from the same panel.

The app contains only synthetic dataset text and requires no secret tokens. It does not handle confidential data.