Extending ContextGuard

LLM providers

• Implement LLMProviderBase.complete_json(prompt, schema, temperature) and pass to LLMJudge.
• Decorate with BudgetedProvider (budgets) and RetryingProvider (backoff/logging) as needed.

Retrievers

• Implement Retriever.search(query, filters, k) returning Chunk with populated provenance, entity_ids, year, optional metadata.doc_type.
• Reuse CanonicalFilters.from_state_spec(state) to translate constraints to your backend.
• For template-method convenience, subclass RetrieverBase (see LangChain/LlamaIndex/Chroma/Qdrant adapters).

Stores

• Implement Store protocol/abstract base for state, facts, runs, and traces using your database of choice.

Numeric/units

• Extend verify.numeric with domain-specific units/scales; add to _UNIT_TOKENS or add new normalization functions.

Profiles

• Add/adjust domain profiles by extending GatingConfig.from_profile and AggregationConfig.from_profile.

Generation

• Implement Generator or subclass LLMGenerator to add streaming, richer schema, or guardrails.