Ensure that a plaintext with k < ciphertext doesn't reduce the ciphertext k. See if replacing this by the noise variance makes more sens or simply remove it (since the maximum precision can be derived from limbs * base2k).