A rubric converts a prompt into weighted, testable criteria for judging generated responses. It enables reinforcement learning when answers lack directly verifiable outcomes, guiding updates toward stronger responses. section 12.6
Museum-exhibit prompt
For a request to suggest five obscure artifacts, hard rules check the count, cultural and temporal distinctness, historical significance, mysterious origins, and thematic intrigue. Principles assess organization, precision, engaging language, and thoughtful selection. section 12.6
{
“initial_reasoning": “<brief rationale for what matters for
this prompt>",
“rubrics": [
{
“reasoning": “<why this criterion matters>",
“criterion": “<clear, testable criterion>",
“weight": <integer 1-10>
},
...
]
}
A rubric configuration records initial reasoning, criterion-specific reasoning, clear testable criteria, and integer weights from 1 to 10. section 12.6
