Self-Critique Validation
Model-checked outputs.
The Model as Its Own Critic
Self-critique uses an LLM to evaluate an LLM's output against a rubric or policy. It catches nuanced failures that deterministic validators cannot encode: factual inconsistency, tone, helpfulness, subtle policy violations.
It is the model-based complement to schema and rule validators.
Separate the Critic from the Author
Run critique as a distinct call with its own prompt, not as a tail instruction in the generation. A clean-context critic, asked only to judge, is far more reliable than asking the author to grade itself mid-generation, where it is biased toward its own answer.
draft = author_model(task_prompt)
verdict = critic_model(
'You are a strict reviewer. Judge ONLY the answer below against the rubric.\n'
'Rubric: ' + rubric + '\nAnswer: ' + draft
)All lessons in this course
- What Are Guardrails
- Input and Output Filtering
- Schema and Rule Validators
- Self-Critique Validation