Replaces hardcoded RAG thresholds with explicit calibrated decision points. Five primitives (retrieval stopping, chunk splitting, context selection, answer abstention, cache trust) behind one swappable state → Decision → confidence → action interface, each evaluated on real datasets with a calibration harness that reports honestly.
jevrag/decision.py — the `ask(state, questions) -> DecisionResult` method calls `api.typesafe.ai/v1/systemone` via the Jev backend. All five primitives in `jevrag/primitives/` route through this method.
Adds structured choices or scores to the workflow; performance and cost benefits have not been independently verified.
The cited decision and RAG-adapter files were read for stub provenance and bypassed relevance values. This site did not rerun a dataset evaluation or paid call. Stub results and None-valued bypasses must remain separate from real Jev results when computing metrics.
Evaluate where learned judgments improve a RAG pipeline instead of replacing every threshold by default.
Material analysed：
decision.py's StubDecision returns fixed values with backend='stub', model='stub' and zero tokens. It exercises measurement plumbing without establishing model quality.
Inspect the supporting source ↗rag_jev_selector.py keeps bypassed relevance as None instead of inventing a score. Coverage and accuracy must preserve these omissions in their denominators.
Inspect the supporting source ↗Useful when you can compare one specific RAG decision against a deterministic baseline. Keep unsupported or low-coverage primitives experimental until held-out evaluation justifies them.
Read DecisionResult metadata and the rag-jev adapter's selection reasons. Recompute metrics with stub/bypassed records separated from actual Jev responses.