As a wholly-owned proprietary product, whatever testing is done on MATLAB is entirely within TMW's internal QA/QC program which they do not publicize. There is no Standards group for MATLAB as there is for the major compiled languages. Although the base product is compiled with such toolset(s), much of the distributed functionality is in the proprietary m-files or other compiled p-code if not compiled completely to native code.
The most that there is you can rely on is that the base math operations are built from such compiler tools with their compliance to the appicable Standard (although surprisingly little of even those Standards has to do with accuracy or correctness of results; they deal mostly with the required semantics and behavior of the compiler and compliance of syntax, etc.) and the references to the available external libraries used like FFTW, LINPACK, etc., for other areas.
If one is in an environment that requires some sort of certification of toolsets, about the only way to do so with MATLAB would be to have an internal test suite against which to compare results. There's at least one frequent contributor here who is in such an environment and does have such a process in place; consequently his organization is relatively slow in accepting new releases. He probably will see this and have more in-depth comments on specifics.
In former life where was involved in licensing and safety-related calculations for power reactors under the NRC and US Code 10CFR40, all calculational tools we used for such calculations had to undergo a verification and acceptance process; MATLAB was not an accepted toolset for such reasons that it would have required another independent verification of any models/calculations done using it to have them accepted for licensing/safety calculations
The vendor- and commercially-supplied Fortran compilers were accepted/used to build the verfied/accepted calculational models referenced above -- any compiler updates required a revisiting of a test suite of those codes to verify the produced the same results when rebuilt with a new compiler version--the compilers themselves were not the verified tool. That would be the same process with MATLAB -- it would be the model that would be verified, not MATLAB itself and reverification of the model with updates to MATLAB. Given the frequency with which TMW releases updates, that could be a sizable burden on an organization to keep up with.
