Concepts¶
Themis has one public mental model now: benchmark authoring on the write side, projection-backed analysis on the read side.
Core Split¶
| Layer | Main types |
|---|---|
| Authoring | ProjectSpec, BenchmarkSpec, SliceSpec, PromptVariantSpec, ParseSpec, ScoreSpec |
| Runtime extension | DatasetProvider, InferenceEngine, Extractor, Metric, JudgeService, PipelineHook |
| Read side | BenchmarkResult, RecordTimelineView, themis-quickcheck |
Start Here¶
- Architecture for the end-to-end flow
- Specs and Records for the public object model
- Plugins and Hooks for extension boundaries
- Storage and Resume for persistence and reuse
- Statistical Comparisons for aggregation semantics