How does Premiere's automatic reverse-telecine work?

I can't seem to find any official technical explanation as to how Premiere determines what the interlaced pulldown cadence of a 29.97 clip is when you edit it into a 23.98 sequence. Does it work by scanning the first few frames of the clip, guessing the cadence (like how After Effects does it in its Interpret Footage panel), and just blindly using that cadence to remove the pulldown on the rest of the clip?

I'm having an issue prepping episodes of a TV show that were delivered to us as interlaced 29.97 DNxHD .mxf files (with interlaced pulldown), and for some reason, some of the episode masters appear to have broken pulldown cadences, usually when a new "act" occurs in the show. I'm not sure what workflow they're using to generate these full 29.97i (59.94i) episode masters, but it causes problems when we try to edit clips from these masters into a 23.98 Premiere timeline, because parts of the master will play back with massive frame-skipping in the 23.98 sequence because Premiere is applying the wrong cadence for the automatic pulldown removal. And yet another part of the same episode file will play back perfectly with removed pulldown.

This wouldn't be a huge problem if Premiere allowed you to force specific pulldown cadences on a clip (similar to the various "WS" pulldown options in After Effects). That way, I could import a second instance of the same clip, and apply a different cadence to it in order to compensate for the broken pulldown cadence in the actual file. But sadly it appears that  there's no way in Premiere to do this.

The interesting thing is that on one specific episode, Premiere's pulldown removal worked perfectly on a scene towards the back end of the clip, but the removal on the entire front end of the episode looked totally janky. My only guess is that because the head of the episode file had a 2-second fade-up from black, Premiere didnt have enough image data on those frames to be able to properly guess the interlaced pulldown cadence?

Right now my workaround for all of this has been to duplicate the episode clip in the Project, create a single-angle Multicam clip for each instance, then edit the in/out points of each act inside of it's own Multicam sequence. Then when you edit the Multicam clip into a 23.98 sequence, Premiere will (apparantly) detect the pulldown cadence for just that section of the clip, and removes it accordingly. 

The one flaw in this workaround is that if you Flatten the Multicam clips in the 23.98 sequence, the "bad" reverse-telecine of the original clip will re-appear. 

Now that Adobe is full-bore into AI, it would be nice for there to be an option to intelligently scan a source file for broken pulldown cadence, and dynamically removes that pulldown on a shot by shot basis. Or at the very least, add the same "Interpret Footage" options that After Effects has (DaVinci Resolve also has a similar option), where you can explictly force a pulldown removal cadence on a clip instance, rather than relying solely on Premiere to do the guessing for you, because clearly it guesses wrong sometimes.