Manual splitting can be inaccurate. Re-timing a run from the proof video allows better comaprison. Here's how I do it, will update this according to feedback.

1. Read all of this enumeration. Agree silently with me, otherwise please debate. I'd like to know a well-defined and reproducible method. :-)
2. Ignore the speedrun timer shown in the video, instead open the run in a tool that shows frame numbers or milliseconds. At 60 fps, there will be 16 or 17 millisecond values that all refer to the same frame. Let's always choose the earliest millisecond value per frame.
3. Find the value A = the milliseconds of the last fully-lit screen of the main menu (multi-episode run) or the last fully-lit screen of the episode selection (single-episode run). To do that, let A' be the earliest millisecond value of the first fading-out screen, then compute A = (A' − 33 ms) for a 30 fps video or A = (A' − 17 ms) for a 60 fps video.
4. Find the value of frame B = the first completely-black frame after the final boss fight. Let B be the earliest millisecond value of this frame.
5. Compute the run's duration as B − A.
6. As a moderator, edit the run and set its time to B − A. I find it good style to add your computations to the run's description: The ignored speedrun timer's value, A, B, and B − A.

It's possible that two players have different loading times. This exact computation from video frames will not solve that problem; it will preserve the difference. You'll need another method or more reasearch to strip loading times.

In 2017, I've re-timed the episode 4 runs tonight according to this. Vortale, CapnClever, and I were nearly tied at 3:45.xx, so I wanted a good, reproducible re-timing method.

From experience, the re-timed times are 0.1 to 0.2 seconds better than the manually-split times because all runners like to split at the end on a guaranteed-black frame, not necessarily on the first black frame.

-- Simon