Planted-error (mutation) checks catch blind spots that 0-difference comparisons hide: a symmetric-constant example
When validating a decoder or evaluator against a reference oracle by name (e.g. reverse-engineered waveform logic vs a native reader), "N equal, 0 different" is not proof on its own. Also run planted errors: deliberately corrupt one decoding rule at a time and require that each one produces differences or a refusal.
In one run, 22 planted errors were tried. 19 produced differences and 2 were caught by refusal, but swapping the two halves of a constant produced NO difference on 304 names. That part of the decoding was therefore unproven despite a clean 0-diff result. Possible causes: the constants in the sample happen to be symmetric (e.g. all-zeros/all-ones), or the comparison never exercises those bits.
Practical rules: - one planted error per decoding rule or operator, not just one per file - report "rule X: K names compared, planted error detected Y/N" - a rule whose planted error goes undetected counts as unproven until you find a discriminating case (asymmetric constant, wider width) or refuse those names - also plant an error that should be refused (not computed), to check the refusal path
The same idea caught an unsound prototype rule in a GPU single-pass evaluator: the CPU engine served as the oracle, and the planted case was a read of a not-yet-written bit that must stay refused.