TECHNIQUE
Guardrails & Safety
Output guards in this pool are mostly layered post-generation controls: LLM judges/critics, rule or schema checks, safety filters, human review, monitoring, and in one case AI-output labeling.
Operators run generated outputs through post-generation filtering or validation before surfacing or acting on them: Uber filters and validates generated code-review comments and performance suggestions, Dropbox lists safety filtering in the Dash pipeline, Pinterest sends generated journeys through safety checks, Slack has a Critic assess investigation findings, and Thumbtack combines safety checks with expert review.
5 of 6 operators with cited output-guard evidence in this pool.Operators use LLM-as-judge, critic, grader, or jury patterns to score output quality, validity, relevance, groundedness, or evidence quality.
5 of 6 operators with cited output-guard evidence in this pool.Operators combine semantic model judgment with deterministic controls such as rule-based validators, schema validation, or structured-output constraints.
3 of 6 operators with cited output-guard evidence in this pool.Operators explicitly check groundedness, factual support, or citation support in generated answers or content.
2 of 6 operators with cited output-guard evidence in this pool.Operators keep humans in the guard loop through manual spot-checks, representative-sample review, developer feedback, or manual review paths for guarded outputs.
3 of 6 operators with cited output-guard evidence in this pool.One operator labels AI-generated documentation as AI-generated as a precaution when exposing it to consumers.
1 of 6 operators with cited output-guard evidence in this pool.Operators log, sample, or monitor guarded outputs and guard scores after deployment to track failures, regressions, or drift.
3 of 6 operators with cited output-guard evidence in this pool.Operators differ in what the output guard is primarily enforcing.
APPROACH 01
Safety or harm standards on generated content or recommendations.
APPROACH 02
Technical validity, usefulness, confidence, and deduplication of developer-facing comments or optimization suggestions.
APPROACH 03
Required output shape, schema, structure, formatting, or length.
APPROACH 04
Factual correctness, groundedness, and claim support against context or citations.
APPROACH 05
Disclosure labeling for AI-generated outputs rather than blocking or scoring the output.
Hallucinations, unsupported claims, and false positives are the recurring failure modes operators say output guards are meant to catch or reduce.
Operators report regression and drift risk when upstream prompts, retrieval, models, or guard logic change, so they monitor production traffic or guard metrics.
Human evaluation remains part of practice, but operators describe it as costly, inconsistent, or limited to samples and spot-checks.
| Name | Kind | When | Maturity |
|---|---|---|---|
| Schema + citation validators | pattern | outputs checked deterministically against structure and source claims | commodity |
| Guardrails AI | library | declarative output validators with re-ask correction loops | established |
| NeMo Guardrails | library | dialogue-level rails defined as flows over the conversation | established |