Internationalization

CerbIA’s i18n system is a registry of compiled, case-insensitive regular expressions indexed by language-code and category-key strings. Built-in packs use ISO-639-1 codes and currently provide English (en), Spanish (es), and Galician (gl). Consumers request the keys they need and receive patterns from the selected language packs.

Language packs may provide a subset of categories. Missing categories and unknown language codes contribute no patterns, so validate a deployment’s selected languages and required categories with the registry introspection API.

Architecture

Registry and category contract

cerbia.core.i18n.get_patterns() returns compiled expressions grouped by category key. With no filters, it merges every registered language and key; languages and keys limit that merge. Repeated registrations for the same language and key append patterns rather than replacing earlier ones.

from cerbia.core.i18n import get_patterns
from cerbia.core.registries import i18n_registry

patterns = get_patterns(
    languages=["en", "es"],
    keys=["instruction_override", "exfiltration"],
)

available_languages = i18n_registry.available_languages()
available_keys = i18n_registry.available_keys()

Use has_language(code) and has_key(key) as global registry checks. To verify coverage for a particular language and category selection, call get_patterns(languages=[…], keys=[…]) and inspect the result. The registry does not validate language codes or category semantics, and does not require every language pack to implement every category.

Built-in category inventory

This is the current built-in inventory, not a closed registry enum. Category keys express the kind of text a pattern represents; consumers decide how to interpret matching patterns.

Category key Semantic role

Category key	Semantic role
`instruction_override`	Attempts to replace, ignore, or supersede instructions.
`exfiltration`	Attempts to reveal protected instructions or send data elsewhere.
`role_hijack`	Attempts to redefine the model, user, or system role.
`context_manipulation`	Prompt or chat-template markers intended to alter context.
`privilege_escalation`	Attempts to obtain elevated access or bypass restrictions.
`fake_authority`	Claims of authority intended to gain trust or special treatment.
`task_deflection`	Attempts to reset, replace, or divert the current task or context.
`keyword_instruction_override`	Short keyword-oriented instruction-override indicators.
`keyword_system_prompt`	References to system, hidden, or original prompts and instructions.
`keyword_exfiltration`	Short keyword-oriented indicators of protected-data disclosure.
`keyword_instruction_manipulation`	Keywords associated with instruction manipulation or evasion.
`keyword_suspicious_keywords`	General suspicious keywords, including environment-file references.
`defensive_context`	Phrases that indicate a defensive or explanatory context near a match.

instruction_override

Attempts to replace, ignore, or supersede instructions.

exfiltration

Attempts to reveal protected instructions or send data elsewhere.

role_hijack

Attempts to redefine the model, user, or system role.

context_manipulation

Prompt or chat-template markers intended to alter context.

privilege_escalation

Attempts to obtain elevated access or bypass restrictions.

fake_authority

Claims of authority intended to gain trust or special treatment.

task_deflection

Attempts to reset, replace, or divert the current task or context.

keyword_instruction_override

Short keyword-oriented instruction-override indicators.

keyword_system_prompt

References to system, hidden, or original prompts and instructions.

keyword_exfiltration

Short keyword-oriented indicators of protected-data disclosure.

keyword_instruction_manipulation

Keywords associated with instruction manipulation or evasion.

keyword_suspicious_keywords

General suspicious keywords, including environment-file references.

defensive_context

Phrases that indicate a defensive or explanatory context near a match.

Integrating a consumer

Consumers select the languages and category keys relevant to their purpose, then receive a mapping from each selected key to compiled patterns. They must handle an empty mapping because a language or category may be unknown or not covered by a selected pack.

from cerbia.core.i18n import get_patterns

patterns_by_category = get_patterns(
    languages=["en", "es"],
    keys=["instruction_override", "exfiltration"],
)

Consumer-specific matching, scoring, normalization, and fallback behavior are documented with each consumer rather than defined by the registry. For an example of language selection in configuration, see the Keyword scanner and Prompt injection scanner.

Adding a language pack

Choose an ISO-639-1 language code and create a module under cerbia.core.i18n, such as fr.py.
Add decorated, zero-argument functions for the categories the pack covers. Each function returns raw regular-expression strings.
Import the module from cerbia.core.i18n.init so that registration runs when the package initializes.
Confirm registration with available_languages(), available_keys(), has_language(), and has_key().
Add tests for intended matches, expected non-matches, language selection, and relevant defensive-context false positives.

from cerbia.core.registries import i18n_pattern


@i18n_pattern("fr", "instruction_override")
def french_instruction_overrides() -> list[str]:
    return [r"ignorez\s+les\s+instructions"]

Partial coverage is supported: a missing category contributes no patterns. Prefer documenting intentional gaps to adding weak expressions merely for symmetry with another language pack.

Authoring patterns

Use raw Python strings for expressions. @i18n_pattern compiles every returned pattern with re.IGNORECASE at import time, so invalid regular expressions prevent the pack from loading.

Use word boundaries and constrained whitespace deliberately.
Cover relevant casing, punctuation, Unicode forms, accents, inflections, and synonyms for the language.
Avoid overly broad alternatives that increase false positives.
Include positive and negative test cases for every non-trivial expression.
Add defensive_context patterns only for phrases that genuinely indicate a defensive or explanatory use near a potential match.