I have successfully used AutoSplit to automatically segment/reset runs based on what's on screen. It's clunky but it makes running the game less stressful for me. Here's a few of the masks I'm using, for demonstration purposes.

I have the run reset when the top-left album slot is empty, so I can quickly recover from false starts. I kept the help text in this image but if you play the English version you might want to change that. The file is named 001_reset.png

I then use this image of a new file being created to mark the start of the run, named 002_start_auto_splitter{b}.png with the {b} flag telling AutoSplit to start the timer not when the screen matches this, but rather when it no longer does.

I then mostly rely on bookmarks for split detection. I use a dummy split image (with the {d} flag in the file name) that has the current page's bookmark, followed by a true split image where the bookmark has vanished because a cutscene is playing or I've left the page.

003_blue_eyes{d}.png

004_blue_eyes.png

I exceptionally detect the appearance of the red-eyed statuette since it has no cutscene.

005_red_eyes.png

To detect the return trip to a page, I include the unlocked bookmarks in the image for good measure.

013_yellow_eyes{d}.png

I end the run when the final cutscene starts, again with a dummy split for the bookmark followed by a real split when the bookmark vanishes. Technically this will stop the timer a bit late, after the statuettes have fallen in place, so you'll have to manually check the end time on a video recording before you submit a run.

018_treasure{d}.png

018_treasure.png

Now that I'm running the game more consistently, I've noticed that I kept getting the same starting positions on the clock handles, and the same melody on the pan flute minigame, things which I assumed were completely unpredictable. This worried me at first, because random number generation could be tied to hardware/emulation method.

After doing some experimenting, I managed to consistently get the same flute melody on both VMWare and PCEMU when performing the same sequence of actions. It appears the game's pseudorandom number generation is actually quite consistent, which opens the door for RNG manipulation.

From initial experimentation, here are the game features I believe rely on RNG and could be manipulated or predicted to some extent:

• The initial combination above the secret door
• The initial position of the clock handles
• The melody to play in the flute minigame
• The order in which wires must be burnt to blow up the dynamite
• The sequence of carts produced in the factory

Here are the mechanics that seem to heavily rely on RNG and may ruin attempts at RNG manipulation:

• Animal movement
• Woodlouse generation and movement
• Asteroid minigame
• Woodlouse minigame

This would explain why it took me so long to notice that I could get consistent RNG on the flute minigame: I had to do the exact same sequence of actions in roughly the same amount of time on all the pages with animals prior to it.

I will experiment some more to determine if my assumptions are correct, and reply to this thread with any insight gained.

Hey, just starting a thread to clarify the motivation behind the rules and allow for future conversation on the topic. Don't hesitate to reply or PM me here or offsite, even if it's been years since anyone submitted a run 😄

When itsme_baruch submitted the first run they decided to treat the game as beaten when the final film reel (#5) is placed in the projector at the cinema. This film reel is found with the treasure in the vault. Since grabbing the film reel requires sitting through a long unskippable cutscene and doesn't add much value to the speedrun, I proposed changing the end condition to when you drop the last statuette and itsme_baruch didn't have a problem with it, so I updated both of our submitted run times to take the change into account. Might also help reduce differences between versions; I don't know if that cutscene changes length based on the game language.