Transpose the random query matrix into independent labels. Project onto the satisfying words, preserving accepted answers and shrinking the decoding set into the original decoding set intersected with those words. Every bad accepting run then supplies a whole fiber of bad evaluation points.