Datasets:
Run commands
run_from_config.sh is the generic YAML entry point:
./run/run_from_config.sh run/configs/project.yaml --dry-run
Install the small YAML-only run dependency when setting up a fresh environment:
python3 -m pip install -r requirements/run.txt
The YAML declares the catalog selection, models, output directory, and an
explicit production executor command. The command is executed once per model
with SBH_MODEL_ID, SBH_RUN_CONFIG, SBH_SCENARIO_COUNT, and the validated
SBH_IDS_FILE selection in its environment. There is no built-in fallback
executor, and an existing receipt is never re-read as if it were a fresh score.
The two-candidate command is deliberately strict:
export CANDIDATE_A_ARTIFACT=/path/to/first/artifact
export CANDIDATE_B_ARTIFACT=/path/to/second/artifact
./run/run_94_example.sh --dry-run
It requires exactly the complete active catalog (currently 94 IDs) and the
two model IDs candidate-a and candidate-b. A real run uses the
run_production_executor.py adapter and requires both artifact directories.
Missing or unresolved paths fail closed instead of claiming that the 94-point
run completed. The executor drives each scenario directly; it never reads old
result rows or verdicts.