The Detection-Extraction Gap: Models Know the Answer Before They Can Say It