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/