Would it be possible to have the uploader check the audio content of files to meet certain thresholds in order to apply these tags?
I think there's a lot of posts with sound_warning that don't really need it, while others have the opposite issue.
To go through them all and check is something that could easily be done by a bot very quickly, but is a tremendous chore for humans.
Aside from standard peak and RMS values, there are some other more sophisticated weightings that could be applied depending on how you'd want it to function.
My reference would be the free Orban Loudness Meter, though I don't know if it could be integrated and/or run in offline mode rather than live to process files quickly.
CBS, ITU BS.1770 Momentary, Short-Term and Integrated, VU, PPM and Reconstructed Peaks are all measured in Orban.
https://www.orban.com/meter
To take 2 example posts, we have this one without a sound_warning tag: https://e621.net/posts/1874774
You've got maybe -6dBFS peaks and maybe -18 more averaged
Then this post, that does have a sound_warning tag: https://e621.net/posts/3041627
This one doesn't even show up on the scale unless I boost the signal massively. (The scales start at -30 and -41)
Not to mention the very slow crescendo in that post, the sound levels don't hit you unexpectedly or anything.
Maybe this is more processing than e6 would consider doing...
Updated