SeedCard Finder 1.0.2: Bedrock End Cities and Structure Fixes

SeedCard Finder is now version 1.0.2, with a corrected Bedrock model. We compared the Bedrock 26.50 model with Minecraft Bedrock 1.26.52 on nine seeds, for 17 kinds of structure. The 1.0.1 model listed 2,205 structures there that the game doesn’t generate, and 1,886 of them were End cities: about two in three End cities it listed were not in the game. Version 1.0.2 lists 35 structures the game doesn’t generate, and none of them is an End city. It also finds more of the real ones. Of the 9,029 structures the game generates at the spots we checked, 1.0.1 found 9,012 and 1.0.2 finds 9,025.

The changes apply to the structure finders, the seed finder and the seed map in Bedrock. Java is not affected. The game’s answers are now stored with the site’s code, so each later change to the Bedrock model is scored against them.

How we compared the model with the game

For each kind of structure, our model first works out the spots where one could start, then decides which of those spots actually get one. This comparison checks the second step: for each spot, does the model make the same decision as the game?

For nine seeds, 12345, 17108149081942125, -4172144997902289642, -5207005594169711425, -3122719057099512664, 42, -1234567890123, 777777777 and 8675309, we took every spot where our Bedrock 26.50 model tries to place a structure within a set distance of X, Z: 0, 0. For each spot, we asked Minecraft Bedrock 1.26.52 for the nearest structure of that kind. If the game’s answer is within 32 blocks of the spot, we count the structure as generated there; otherwise it isn’t. The distances were 5,000 blocks for End cities, 2,500 to 4,500 Nether blocks for Nether structures, and 700 to 25,000 blocks for Overworld structures, wider for the rarer ones. That is 42,405 spots in all.

We worked out the new rules on the first five seeds. Three more, 42, -1234567890123 and 777777777, were held out from every fit we have done, including earlier ones: we scored the model on them only after the rules were chosen. On those three seeds, 1.0.1 listed 701 structures the game doesn’t generate and 1.0.2 lists 9; 1.0.1 missed 6 real structures and 1.0.2 misses 4. (Bastions on 42 aren’t counted there, because an error on that seed helped suggest the bastion fix described below.) We didn’t use 8675309 to choose these rules, but it helped with an earlier mansion and trail ruin fix, so it only counts towards the nine-seed totals.

We also counted each structure’s own blocks in the generated world. At all 30 spots where 1.0.2 still lists a village, temple, ancient city, ocean monument or trial chamber that the game doesn’t generate, the world had at most 2 of that structure’s blocks. At the 4 real structures 1.0.2 still misses, it had between 451 and 40,293. A buried treasure is a single chest, so we couldn’t count those.

End cities: the ground has to be high enough

The 1.0.1 model listed an End city wherever its spot landed in the End highlands or End midlands. Our comparison showed that the game also checks the ground. It looks at the ground near the middle of the city’s chunk and at three points five blocks from there, in the direction the city is turned. At all four points, the first air block above the ground must be at Y 60 or higher.

Version 1.0.2 applies the same height check. End city positions didn’t change; only which spots are listed did.

The rule holds on the seeds we didn’t use to work it out. On the three held-out seeds, the game generates 354 End cities among the model’s spots. The 1.0.1 model listed all 354 plus 606 more the game doesn’t generate; 1.0.2 lists the 354 and nothing else.

Here is one seed in full: every End city spot within 5,000 blocks of 0, 0 for seed 12345, Bedrock 26.50.

The green dots are End cities the game generates; both versions list them. The orange rings are the 209 spots the 1.0.1 model listed where the game generates no city. Version 1.0.2 drops all of them and lists only the 113 green dots.

Every other structure

“Listed, not generated” counts spots where the model lists a structure and the game generates none. “Missed” counts structures the game generates that the model doesn’t list. Every row covers all nine seeds.

What changed in 1.0.2, apart from End cities:

• Ocean monuments need deep ocean all around the monument, and lush caves under the sea floor now rule one out.
• Shipwrecks, ocean ruins and buried treasure are checked at the height of the sea floor, so a cave biome below the sea floor rules them out.
• Villages and pillager outposts are checked at the surface, so a cave biome just under the surface rules them out.
• The three strongholds placed near villages follow the village rule and Bedrock’s own direction calculation.
• Woodland mansions need dark forest (or pale garden) everywhere within 32 blocks of the middle of their chunk, checked from the top of the world so cave biomes don’t count.
• Ancient cities need the deep dark at the corner of their chunk, at Y -27. The 1.0.1 model checked the middle of the chunk.
• Bastion remnants check the corner of their chunk and the three neighbouring spots to the west, north and north-west; none can be basalt deltas. Two bastion errors on seeds 42 and 8675309, outside the first five, first suggested this wider check. We only adopted it after it predicted six bastions the game doesn’t generate, in areas beyond 2,500 blocks on the first five seeds that we recorded afterwards; the single-corner check from 1.0.1 lists all six.
• Trial chambers can’t start in sulfur caves.
• Trail ruins check their biome at the corner of the chunk, 15 blocks below the surface.
• Temples, ruined portals and Nether fortresses keep their checks. Ruined portals and fortresses already matched the game at every spot.

What’s still off: 12 villages, 8 temples, 5 ancient cities, 5 buried treasures, 3 ocean monuments and 2 trial chambers are listed but not generated, and 3 trail ruins and 1 trial chamber are missed. The remaining villages and temples sit in the same biomes and above the same caves as real ones, so we haven’t found a rule that tells them apart yet.

Desert pyramids, jungle temples, witch huts and igloos are counted together. The game looks them up as one kind of structure, so a spot counts as listed when the model lists any of the four there. Our comparison doesn’t say which of the four the game chose.

Separately, Bedrock maps used to show grove, snowy slopes, jagged peaks and frozen peaks as unknown biomes. They now have their names.

Every model change is now tested against the game

We keep the game’s answers for all of these spots with the site’s code. Each new build of the Bedrock model is scored against them, structure by structure. If a build lists more structures the game doesn’t generate, or misses more that it does, our tests fail. A later fix for one structure can’t quietly make another one worse on these seeds.

What this comparison covers

• Where structures start, in Minecraft Bedrock 1.26.52 and our Bedrock 26.50 model. It doesn’t cover structure pieces such as End ships, loot, or the terrain around a structure.
• Only the nine seeds and the areas around 0, 0 described above. For every other seed and place, the finders and seed map show model predictions.
• Not abandoned camps, Nether fossils or dried ghasts, which the game can’t look up.
• Bedrock only. Our Java models are separate and were not changed.

You can search for Bedrock seeds in the seed finder, find the nearest End city with the End city finder, or open your own seed on the seed map. Our earlier SeedCard Finder 1.0.1 article covers the first Bedrock bastion fix, and the About page explains what a model prediction can and can’t tell you.