Meridian Line wants an automated check that screens every generated response from the safety-review summarizer for unsafe, harassing, or otherwise policy-violating content before it reaches a human reviewer, without hand-writing a list of banned phrases. What kind of guardrail component fits?

An output guardrail implemented as a safety classifier model scores generated text for categories like harassment or unsafe content and blocks or flags anything above a chosen threshold, which generalizes far better than a fixed keyword list because it reasons about meaning and phrasing rather than requiring every problematic term to be enumerated in advance -- this is the standard applied-guardrail pattern for automated content screening. A regex banned-word filter is brittle: it misses rephrasing, misspellings, and context-dependent harm, and produces both false negatives (unsafe content that avoids the listed words) and false positives (safe content that happens to contain a listed word), so it does not meet the 'without hand-writing a list' requirement even where it partially works. Asking the model to 'be safe and appropriate' in the system prompt is a soft instruction with no verification step -- the model can still fail to follow it, and there is no scoring or blocking mechanism enforcing the outcome, so it is not a guardrail in the applied-guardrails sense at all. Manual review of every single response does not scale and is explicitly the workload the automated guardrail is meant to reduce, not the guardrail itself. Exam framing: implementing LLM guardrails on output is a named Section 3 objective, and a safety-classifier-based screen is the concrete mechanism most often meant by it.