Skip to content

Glossary

  • artifact: persisted payload produced by generation or evaluation and stored for later inspection
  • benchmark: named catalog entry that resolves to a dataset plus any adapter-specific behavior
  • candidate: one generation attempt for a case
  • candidate set subject: a workflow-backed evaluation subject containing one or more candidates
  • case: one dataset row evaluated by the runtime
  • component: a generator, reducer, parser, metric, or judge model with component_id, version, and fingerprint()
  • explanation: a concept document that helps the reader understand why Themis works the way it does
  • how-to guide: a task-oriented document that helps an already-motivated user achieve a specific outcome
  • provenance: metadata recorded with a run that does not change run_id
  • reduced candidate: the candidate selected or synthesized after reduction
  • reference: lookup-oriented documentation for exact commands, types, fields, and interfaces
  • RunSnapshot: the immutable compiled artifact that captures identity, provenance, datasets, and resolved component refs
  • run: one execution of a compiled snapshot
  • score: the final metric output recorded for a metric on a case
  • subject: the object passed into workflow-backed scoring, such as a candidate set, trace, or conversation
  • trace: structured step-by-step execution data captured from generation or evaluation
  • tutorial: a learning-oriented lesson that guides a newcomer to a successful outcome