The resolution of binary signals in Middleton's class A narrowband non-Gaussian noise for both coherent and incoherent threshold or weak-signal detection is treated. A specific relative crosscorrelation coefficient p is defined for narrowband binary signals, and it is then shown that: (i) for coherent detection, `antipodal?? (p = ??1) signals are best, i.e. they give the smallest error probability Pe in the threshold regime, and (ii) for incoherent detection, `orthogonal?? (p = 0) signals are best. Threshold performance, i.e. Pe, is a monotonic function of NSL, where N is the number of large (independent) data samples, S is the normalised signal-to-noise ratio and L is Fisher's information measure for the noise statistics.