Operations & Hardening

Budgets and guardrails

• Retrieval: clamp claims and total_k; enforce MAX_JUDGE_CHUNKS_PER_CLAIM and MAX_JUDGE_TEXT_LEN.
• LLM: enforce prompt/output budgets with BudgetedProvider; wrap with RetryingProvider for retry/backoff/jitter.
• Planner/gate/judge respect domain profiles for stricter defaults.

Logging/metrics

• Async runner: pass logger or Instrumentation to async_run_verification; emits events for plan/retrieve/gate/judge/aggregate plus errors.
• Wrap providers with your own logger or subclass RetryingProvider/BudgetedProvider to emit metrics.
• Trace: use TraceGraph/TraceBuilder to capture every decision; export DOT for audits.

Error handling

• Retrieval/judge exceptions in async runner are logged (if logger provided) and re-raised.
• Circuit-breaker wrappers available: CircuitBreakerProvider, CircuitBreakerRetriever; RetryingProvider/RetryingRetriever for retry/backoff.
• Keep provider wrappers thin; layer retries/backoff externally where possible.

Security notes

• Prompt hardening in LLMJudge prompts; still apply your own red-team/guardrails at provider edges.
• Enforce source policy via SourcePolicy and gating; populate provenance to avoid silent acceptance.

Performance

• Use async runner to parallelize retrieval + judge; limit concurrency via max_concurrent_tasks.
• Use backend-side filtering via CanonicalFilters to reduce post-filter load; optional async retriever avoids threadpools.
• Cache retrieval if backend supports it (see RetrieverBase cache flag); dedup/rerank with retrieve.rerank helpers.