Coverage Failure as a Diagnostic Signal in Neural Model Probing

We report a methodological observation from a multi-machine dense-grid activation probing study of nineteen small to mid-sized language models (124M to 7.6B parameters). Every experiment satisfied a nominal threshold of one hundred million observation slots, yet observed fill rate varied from 58.75 percent to 97.05 percent across experiments. The observed gap is strongly associated with protocol choices, in particular the prompt corpus size and the use of per-channel quantile binning. We cannot yet rule out interaction with the network under study, the prompt set, hook discovery completeness, or quantile edge quality; the comparisons we report are observational rather than controlled. We argue that total slot count alone is an invalid completion criterion for activation probing, that coverage percentage should be reported as a first-class result, and that under-coverage should be treated as a diagnostic signal about the probe protocol before being attributed to the model.