On Evidence Capture for Accountable AI Systems