Metric families and subjects¶
What it is: the relationship between metric families and the subjects they consume.
When it matters: whenever you are choosing a metric type or trying to understand why a workflow expects a candidate set, trace, or conversation instead of just parsed output.
What you provide:
- pure metrics consume parsed output and case data
- LLM metrics consume candidate-set subjects
- selection metrics consume candidate-set subjects for pairwise or ranked comparisons
- trace metrics consume trace or conversation subjects
What Themis provides: subject construction, workflow execution, persistence, and artifact inspection.
Use this map when the metric family seems right but the evidence shape does not.
flowchart TD
A["Themis subject builder"] --> B["Parsed output + case data"]
A --> C["Candidate set"]
A --> D["Trace or conversation"]
B --> E["PureMetric"]
C --> F["LLMMetric"]
C --> G["SelectionMetric"]
D --> H["TraceMetric"]
Metric families are mostly distinguished by the subject shape they need, not just by how they compute the final score.
What to inspect when it goes wrong: verify that the subject type and metric family match the kind of evidence you want the runtime to score.