Skip to content

Output Validation

The Output Validation module runs 7 parallel quality checks on LLM-generated output. Each check returns a decision; the engine aggregates to the most restrictive.

Checks

Check Detects Decision on Fail
Schema Validity Invalid JSON or schema mismatch repair
Citation Check Missing citations when required repair
Hallucination Proxy Claims not found in context repair or reject
Policy Check Disclosure patterns, policy violations repair
Unsafe Language Harmful instructions in output reject
Confidence Threshold Excessive hedging language escalate
Genericity Detector Generic filler phrases ("AI slop") repair

Decision Priority

reject > escalate > repair > pass

Endpoint

POST /v1/outputs/validate

Source

src/agentguard/output_validation/