Hudson's paper, ''Novelty and the 1919 Eclipse Experiments,'' focuses on two main questions, both of which have been the subject of debate among philosophers of science: The first is (1) when, and why, should novel data be preferred or required in evaluating hypotheses and theories? The second is (2) were there good grounds for discounting one of the eclipse results in the 1919 testing of Einstein's General Theory of Relativity? Hudson sets out to provide answers to (1) and (2) that differ from, or improve upon, those that I discuss (in Mayo, 1991(in Mayo, , 1996)), but I do not see that he has made his case. He has not shown that his answer to (1) goes beyond a rewording of my condition; nor has he adduced historical information that speaks to an altered answer to question (2) (i.e., a ''no'' rather than a ''yes''). Nevertheless, further discussion of these issues is needed, and the main goal of my comments will be to highlight two broad problem areas or queries to which new experimentalists might turn their attention-if they are seeking to make real and interesting progress beyond what has already been done. The first is (i) when do violations of use-novelty, and more generally, ''double counting'' of data prevent hypotheses from passing severe or reliable tests? The second is (ii) what are some general strategies for informal assessments of reliability or severity? Can we identify a handful of canonical models of error and informal arguments from error? Tackling these, it seems to me, will require careful study of historical episodes including an analysis-or reanalysis-of the relevant data and statistical techniques.
1. Severity as the rationale for the requirement of use-novelty
Hudson claims he will ''reformulate use-novelty in a way that accommodates Mayo's insight'' (in Mayo, 1996) that hypotheses can be use-constructed by means of highly reliable use-construction rules. Some examples I give there of reliable