I studied the statistics of detected neutrons that leaked from four subcritical re flected, enriched-uranium assemblies, to explore the feasibility of developing a crit icality warning system based on neutron noise analysis. The calculated multiplica tion factors of the assemblies are 0.59, 0.74, 0.82, and 0.92.1 studied three possible discriminators, i.e., three signatures that might be used to discriminate among as semblies of various multiplications. They are (1) variance-to-mean ratio of the counts in a time bin (VIM), (2) covariance-to-mean ratio of the counts in a common time bin from two different detectors (C/M), (3) covariance-to-mean ratio of the counts from a single detector in two adjacent time bins of equal length, which I call the serial-covariance-to-mean ratio (SC/M). The performances of the three discrimi nators were not greatly different, but a hierarchy did emerge: SC/M ^ V/M ^ C/M. An example of some results: in the neighborhood of k = 0.6 the Afc required for satisfactory discrimination varies from about 3% to 7% as detector solid angle varies from 19% to 5%. In the neighborhood oik = 0.8 the corresponding Ms are 1% and 2%. The noise analysis techniques studied performed well enough in deeply subcritical situations to deserve testing in an applications environment. They have a good chance of detecting changes in reactivity that are potentially dangerous. One can expect sharpest results when doing comparisons, i.e., when comparing two records, one taken in the past under circumstances known to be normal and one taken now to search for change.