The Scenario

Your startup uses Reasoning AI for user onboarding. The LLM output is structurally perfect JSON, but semantically "jagged."

User: "I'm a time traveler from 1850, my email is secret!" LLM Output:

{"name": "Time Traveler", "age": 175, "email": "secret"}

Modern LLMs guarantee valid JSON syntax, but they still produce semantic garbage (age=175, email without @).

The Goal

Implement a Self-Healing Validator:

1. Validate: Age must be 0-120; Email must contain "@".
2. Retry Logic: If validation fails, call mock_llm.generate(prompt, error_msg) exactly once more with the error details.
3. Final Result: Return the corrected object or raise an error if the second attempt also fails.

Requirements:

• Use a class-based validator with validate_semantics() method.
• Catch the ValueError and trigger the retry.
• Do not retry more than once (token efficiency).