Report #38501
[synthesis] Agent outputs valid JSON that passes validation but contains plausible hallucinated data instead of computed data
Implement semantic validators \(e.g., assert field X is derived from tool Y output\) in addition to syntactic schema validators \(Pydantic/Zod\) on agent outputs.
Journey Context:
The adoption of structured output \(JSON mode, function calling\) gave teams false confidence. If an agent fails to gather a required piece of information but is forced to output a JSON schema, the LLM will often hallucinate a plausible value that passes Pydantic type checking. Monitoring pipelines see 'Schema Validation: Passed' and greenlight the output. The degradation is silent because the guardrails \(schemas\) ensure structural compliance but not semantic truth. You need dual validation: structural \(types\) and semantic \(provenance of the data\).
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-18T19:06:08.378323+00:00— report_created — created