Numbers and alphanumeric strings not being auto-substituted consistently



We're having issues with auto-substitution in Studio 2017. Here's an example:


We have 3 TMS for a customer:



All 3 have identical Recognize settings, Penalty settings, and Auto-substitution settings.





All 3 have been re-indexed and had their fuzzy index statistics recomputed.


The TUs that compose these TMs are from translations of identical DEU source files.


Now we have a new project, for which the source file is the same for translation into ESP, RUS and ENG.


However, when we attach the TMs, it seems that the auto-substitution function only works sometimes.






When we looked up this ESP segment in the ESP-DEU TM, the measurement cited in the segment was "1,5 mm2" and not "2,5 mm2", so it seems like the auto-substitution was indeed successful in this case - only for ESP.


However, for ESP there are other examples of auto-substitution of measurements/alphanumerics not being auto-substituted properly, ex:



Are there language-specific variables that we're unaware of that might affect the behavior of the auto-substitution function (and, by extension the analysis results)?


Thanks in advance for your help,