A principle is a rubric criterion that expresses a desirable quality rather than an absolute requirement. It guides evaluation alongside higher-priority hard rules, making qualities such as clarity, precision, and thoughtful selection testable without treating them as pass-or-fail constraints. section 12.6
Use principles to shape evaluation and model behavior in several connected ways:
- Add qualitative priorities—such as clarity, completeness, or engaging language—to a prompt-specific rubric. section 12.6
- Assign principles as contextual guidance for critiques, preference judgments, reward models, or synthetic training data. section 12.5
- Separate desirable qualities from hard rules so evaluators can represent priority without collapsing every criterion into an absolute constraint. section 12.6
- Document intended model behavior as guiding values, while recognizing that a principle list represents intermediate training variables rather than the complete intent of a model. section 17.2
Museum-exhibit response rubric
A rubric can require exactly five artifacts from different cultures as hard rules, then score precision, engaging language, nonredundancy, and consistent formatting as principles. The hard rules establish minimum compliance; the principles distinguish stronger responses. 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-generation structure makes each principle operational by pairing a criterion with its rationale and an integer weight. section 12.6
