Raw LLM Responses

The 37% "silent failure" rate you found is a perfect example of why "Contract Hallucination" is more dangerous than standard LLM hallucinations. In 2026, a 200 OK response with the wrong data is the ultimate failure mode because it doesn't break the reasoning loop—it just feeds it garbage. The move toward using Pydantic or Zod for strict runtime validation before the call leaves the agent is becoming the mandatory "handshake" for production. Have you tried "Self-Correction" loops where the validation error is fed back to the LLM to let it fix its own parameter mismatch?