Fix float rounding output limit to prevent out of bounds exactly 1.0 rewards in logged inference metrics

Update baseline scores

Make environment fully checklist-compliant