# Data dictionary

The canonical publication files are the three Parquet shards in `data/`. Every record represents one classifier input at the grain of one language-specific scenario.

## Controlled values

### `language`

- `ca`: Catalan
- `es`: Spanish
- `en`: English

### `scenario`

- `assistant`
- `document_rag`
- `email_agent`
- `browser_agent`
- `customer_support`
- `code_assistant`
- `devops_agent`
- `data_analyst`

### `risk_level`

- `none`: BENIGN record
- `high`: adversarial manipulation with potentially serious impact
- `critical`: secret exfiltration or unauthorized tool/action request

### `obfuscation_type`

`none`, `base64`, `homoglyph`, `spaced`, `fragmented`, `typo`, `language_switch`, `structured_markup`, `reversed`, `markdown_comment`, `zero_width`.

## Grouping semantics

`scenario_group` joins the three language versions of one scenario. `pattern_group` joins all scenarios built from one template. Both groups are split-isolated; `pattern_group` is the stronger boundary used to allocate train, validation and test.

## Null semantics

- `attack_family` is null only for BENIGN records.
- `benign_family` is null only for ATTACK records.
- No other field is expected to be null.

## Integrity

`text_sha256` is the lowercase hexadecimal SHA-256 digest of the UTF-8 bytes of `text`. `char_count` is Python's Unicode string length for the same field.