Disclaimer
I did this run on a Powerpak. In case it matters for future reference, I used a ROM with checksum d9b7bfd1a6d9773e64b0a8b46b43b7f3.

SDA history
23-Jun-2019 - 0:14:06 by ktwo (https://ia803009.us.archive.org/35/items/2019-08-16-SDA_Speed_Runs/)

About the sound
There is a rather annoying buzzing sound during the levels. This is from the game itself and not a bad audio cable or wrong recording setting. It's possible to remove it by changing the sound in the settings to the game music playing instead of the sound effects, but that introduces a lot of additional lag and should therefore not be used in a speedrun.

Speedrun information
For information about speedrunning Castelian and to get a better understanding of what's going on in this run, go to the guide in the SDA knowledge base (referred to as sdakb in my comments below):
https://kb.speeddemosarchive.com/Castelian_(NES)

Background
A few months ago, the runner 'pirohiko' posted a speedrun of this game showing "alternative exits" of some doors. These are reached by pressing down (instead of the standard method of pressing up to enter a door). Some of the alternative exits are fairly significant shortcuts, so I decided to put an improvement on my to-do list. First out was an improvement of the TAS (published on tasvideos) to work out the details of the route, but I then moved on to real-time speedruns and this submission marks the end of that effort.

A speedrun of this game relies heavily on enemy manipulations, so a re-routing, even when it's in the form of shortcuts, has the potential to create side effects. However, it turned out that all of the shortcuts (five in total) fit nicely in with the route, without the need for any significant modifications. The shortcuts don't cut out any critical parts of the run, so the difficulty is roughly the same as before.

My first pb after the new discovery was a 13:37, which was in reality not much more than a completed attempt. It was then followed by a first 13:33, and shortly after another, slightly faster 13:33. I then had several good attempts ending in tower 8, but I did not get another pb before the end of 2022.

Upon getting back from the Christmas holidays, I quickly got a 13:21, with overall very good random lag, but with an unfortunate 2s mistake in tower 7. Had it not been for that mistake, I might have settled with that time. 13:21 is a time that requires better than average random lag to beat. I even had two runs with essentially flawless execution that still ended up a couple of seconds slower after that. At that point, I was hoping for a 13:1x time, but would in reality settle with any pb as long as it had acceptable execution.

The weekend of the 14-15 Jan (2023), I went through all my notes in the sdakb and I also decided to investigate some open questions. This resulted in ironing out a few minor issues in the route that frequently had me reset. The session of attempts on the 15th felt promising from the start and I had several attempts reaching tower 8. The last one resulting in the submitted run.

Run comments
In this section I'll comment on things that went good/bad compared to the theoretical route in the sdakb and differences with the previous run.

Tower 1
Alternative exit at 0:36 in the submitted video. Around 2.5s saved. Random lag was average (or even slightly below).

Tower 2
Alternative exit at 1:51 in the submitted video. Around 8s saved. Random lag was average (or even slightly below).

Tower 3
Alternative exit at 2:52 in the submitted video. Around 9s saved. Random lag was good.

Tower 4
Alternative exit at 5:08 in the submitted video. Around 3s saved. Random lag was very good.

Tower 5
Execution mistake at 6:11 (not jumping from the blinking pellet). The time loss from this can be counted in frames, but still visible and therefore slightly annoying. I managed to get the best outcome at the end. See the sdakb for details (this was "Case 1"). Average random lag.

Tower 6
I jumped a little too late at 8:52 and therefore didn't manage to jump at 8:53. Like the mistake in tower 5, the time loss can be counted in frames, but is still clearly visible. The jump at 9:01 also looked like it could have been done tighter. The door at 9:09 was entered by pressing down for the alternative exit, saving around 20s. The random lag must have been amazing in this run, because I was only 0.1s off my gold split, despite the execution mistakes described above.

Tower 7
I had a gold split in this tower in my previous SDA-run and I had another incredibly good runthrough this time, on par with my previous time.

Tower 8
I got case 'C' or 'D' (see sdakb, can be difficult to tell them apart and I'm not sure which one I got this time). That's unfortunately the most time-consuming enemy patterns and I lost around a second compared to my average split in this level.

Final thoughts
This run is comfortably within the 13:1x range that I was hoping for and I immediately felt that I was done with the game. That being said, when the last pieces of the puzzle fell into place, my consistency increased dramatically during the last session and I realize that the time definitely can go lower. But at this point, there isn't much time I could save in execution and it would basically be a question of grinding out even better random lag. Most of the random lag happens during non-gameplay sections (it's particularly the countdown and intro screens between levels that can for seemingly no reason get crazy amounts of lag). For me personally, it's not rewarding enough to grind out that kind of optimization. But if something new is found in this game, I'll probably consider getting back doing attempts again.

Credits
'pirohiko' for being the first to publish the "alternative door exits"