The article evaluates six open-source libraries for implementing LLM guardrails, emphasizing that no single tool addresses all risks and a layered architecture is necessary. Guardrails AI provides a validator framework for input/output validation and structured data checks via Guardrails Hub, suitable for reusable validation pipelines. NVIDIA NeMo Guardrails uses Colang for programmable conversational policies, dialog rails, and tool execution controls across multi-turn interactions. Microsoft Presidio specializes in PII detection and anonymization using rule-based recognizers and NER, supporting local deployment for sensitive data redaction in prompts, retrieval, and responses. Pydantic offers deterministic schema validation for structured outputs, enforcing type constraints, field restrictions, and JSON parsing without additional LLM calls, ideal for API boundaries. LLM Guard (archived July 2026) implements modular scanners for prompt injection, toxicity, and sensitive data detection but lacks ongoing maintenance. Outlines constrains generation to schemas (JSON, Pydantic models) via supported inference backends, reducing malformed outputs before they occur. The article proposes a seven-layer reference architecture: authentication, input validation, PII handling (Presidio), model interaction controls (NeMo), structured output validation (Outlines/Pydantic), tool authorization, and monitoring. Evaluation requires threat modeling, positive/negative test cases, precision/recall metrics, latency tracking, realistic attack testing (including retrieved content), failure-mode analysis, and regression testing. Common mistakes include relying solely on system prompts, conflating schema validity with factual correctness, assuming complete PII detection, over-filtering, ignoring retrieval risks, using unmaintained libraries, delegating authorization to the LLM, and skipping production monitoring.