discovery-by-llm/eval/run_eval.py

23 lines
642 B
Python

"""Eval harness entry point.
Fixture mode (default, key-free): byte-snapshots of deterministic pipeline
stages against recorded cassettes. Live mode: property assertions only —
LLM output is constrained, not pinned.
"""
from pathlib import Path
import yaml
SCENARIOS_FILE = Path(__file__).parent / "scenarios.yaml"
def load_scenarios() -> dict:
"""Load the shared scenario definitions."""
with SCENARIOS_FILE.open() as handle:
return yaml.safe_load(handle)
if __name__ == "__main__":
scenarios = load_scenarios()
print(f"{len(scenarios['scenarios'])} scenarios, {len(scenarios['refinements'])} refinements")