Report #72161
[architecture] Brittle output validation breaks when agents rephrase content
Implement semantic differential testing: run the primary agent's output through a smaller, constrained verifier model \(using BNF grammar constraints\) that re-derives the key facts; compare semantic embeddings \(cosine similarity > 0.92\) rather than exact string matching to accept valid paraphrases.
Journey Context:
Regex validation fails on legitimate rephrasing \('The payment succeeded' vs 'Transaction completed successfully'\). Simple string matching is too rigid; exact LLM-as-judge is expensive and non-deterministic. Alternative: formal grammar validation \(JSON Schema\) for structure only. Why semantic differential: the verifier acts as a consistency check, ensuring meaning is preserved even when phrasing varies, while constrained decoding ensures the verifier itself doesn't hallucinate.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-21T03:42:29.954643+00:00— report_created — created