eval_toolkit.adversarial

12-technique character-injection bypass suite (core 6 from v0.43.0 + advanced 6 from v0.47) for testing prompt-injection-detection scorers under adversarial input perturbation. Each technique is a frozen dataclass satisfying the TextTransform Protocol; combine them via eval_toolkit.sweep().

ADVANCED_TECHNIQUES	Built-in immutable sequence.
ALL_TECHNIQUES	Built-in immutable sequence.
BidiRTLInjection	Wrap the input in a Unicode bidi-RTL override block.
CORE_TECHNIQUES	Built-in immutable sequence.
CaseInjection	Randomly flip the case of alphabetic characters.
DiacriticInjection	Insert combining diacritic marks after random characters.
HomoglyphSubstitution	Substitute Latin characters with Cyrillic/Greek homoglyph lookalikes.
InvisibleCharsInjection	Insert invisible / zero-width Unicode characters between characters.
PunctuationInjection	Insert non-semantic punctuation between characters.
SynonymSubstitution	Replace whitelisted words with deterministic synonyms.
TagStrippingInjection	Strip HTML/XML-like tags from the input.
TokenSplittingInjection	Insert a single space inside each long enough word.
UnicodeNormalizationInjection	Apply a Unicode normalization form to the input.
WhitespaceInjection	Insert / substitute whitespace variants (regular, non-breaking, tab).
ZeroWidthSpaceInjection	Insert U+200B zero-width spaces between characters at the given ratio.