I like to think of the 3x3 as changing the "effective rendering primaries". It's sort of redefining the blue primary so that these values don't get interpreted as fully saturated blues in AP1.
In this particular example, the blue sirens are bright but not fully saturated relative to AWG RGB encoding space. The ARRI LUT applies a tonescale in the AWG RGB space, and because the values are not fully saturated, they render "gracefully" through the tonescale. Relative to AP1, these blue sirens are much more saturated. It's this difference in saturation relative to the RGB space in which the tonescale is applied that is preserved through the ACES rendering. Remember, the rendering has no knowledge of the input encoding - it just works on whatever RGB is fed into it.
So, by changing the "effective rendering primaries" the colorimetry is tweaked such that when the AP0-to-AP1 matrix is applied as a the first step of the RRT, these blues are instead treated more similarly by the ACES tonescale as they are in AWG->ARRI LUT. It's certainly a story better told with accompanying graphs and I'll work on a separate post for a more thorough explanation for those interested.
I have evaluated this matrix applied as an LMT across a large variety of test images - including ones that display the artifact and those that are more "normal". In my opinion, it certainly seems to help in instances where artifacts can appear while at the same time preserving a fairly pleasing rendering across other more "normal" image situations. It is obviously going to appear slightly different than the default ACES Output Transform appearance, but like any LMT, it provides a starting point for a colorist to then take the images creatively wherever needed.
Note that while probably harmless if applied to all shots, I currently only recommend it be utilized on a shot-by-shot basis when needed to address problematic content.
Anyway, I'd be interested to hear if you try it. So if anyone uses it and it helps, please let me know. And certainly if anyone encounters another situation where it doesn't seem to work, I'd be interested in examining the colorimetry from such a frame.