Perhaps there was a glitch in the auditory presentation of stimulus A in this one experiment? Because of the random mixture of events, the subject wouldn't notice (unlike a missing block of 30 s, say, which would be kind of obvious). But how you could verify this, without having a recording of the sounds that the subject actually heard, is something I can't say. I suppose it's too late to debrief the subject and found out if he/she heard all the class A sentences?

One 'trick' you could try, assuming you are using '-stim_times' for 3dDeconvolve modeling, is to use '-stim_times_IM' to compute separate amplitudes for EACH event (saved via '-cbucket'). These will be very noisy, and individually meaningless, but you can then plot them in the auditory regions and see if you find any pattern -- for example, that some class A stimuli DID have large responses, but some didn't, and that the variance in the response magnitudes is so large that no average effect is significant -- this could be the case, for example, if there was some volume glitch in the presentation of class A sentences (sometimes loud, sometimes not). You can even use the newish 'Opt->Colors->Boxes' plot mode to graph these amplitudes, rather than graph them as a time series.

I was able to find an error in some data this way (albeit with a block design, which is simpler to deal with) -- one run had the timing off and all the blocks were out of phase, giving negative amplitudes. By plotting individual block responses, this become blindingly clear (lots of positive responses, with a set of negative responses in the middle). Convincing the originator of the data was a different story, of course.