bintzgavin Apastra baseline
Locate the intended run directory and the adapter that produced it. Confirm the suite and model set with the user when several runs fit their request. A scorecard alone is insufficient.
Run the admission command before accepting evidence:
.agent/bin/apastra gate promptops/runs/candidate --adapter promptops/harnesses/local.yaml
.agent/bin/apastra baseline smoke accepted-001 promptops/runs/candidate --adapter promptops/harnesses/local.yaml
The baseline command revalidates the retained evidence. It creates
derived-index/baselines/smoke-accepted-001.json using exclusive creation and
records the resolved run path and its digest.
Report the path produced by the command and its measured metrics. Preserve existing baseline names. Accept a later run under a new name, then explicitly select its run directory when comparing:
.agent/bin/apastra gate promptops/runs/candidate --baseline promptops/runs/accepted --policy promptops/policies/regression.yaml --adapter promptops/harnesses/local.yaml
A failed run, unverified producer, or unresolved digest cannot establish a baseline. A missing baseline requires an explicit bootstrap choice. Retain historical scorecard-only records without promoting them to current evidence.
For model comparison, use the measured comparison command:
.agent/bin/apastra compare smoke --adapter promptops/harnesses/local.yaml --models model-a model-b
The evaluation contract describes the required files and supported policy semantics.