A tank fighting the boss. A Cleric healing the tank. Damage dealers attacking. On paper, that looks like a party doing its job.
Then another monster reaches the Cleric, and the run starts falling apart.
That pattern appears in our dungeon investigations. It is a useful example of what improving the AI actually involves: finding the moment when several individually reasonable actions produce a terrible group result.
The enemy nobody owned
One diagnostic sample recorded 240 occasions when a Cleric attracted enemy attention. In 123 of them, the new attacker was a monster the damage dealers had not touched. These incidents followed routine healing while the tank was still at roughly 60–80% health.
Those numbers describe that sample, not every Cleric death on the server. They also explain why “heal sooner” is not a complete answer. The healer was already healing. The group needed to deal with the enemy that healing had brought into the fight.
From the Cleric’s point of view, the tank needed help. From a damage dealer’s point of view, the boss was still the target. Between those two views sat a loose monster with nobody taking responsibility for it.
Fix the situation, then test the fix
The work that followed addressed several parts of the same problem: protecting the healer, dealing with extra enemies, and getting separated characters back to their group safely.
That last detail matters more than it sounds. Another recorded failure involved a damage dealer taking a direct route back to its party through a pack of spiders. Being close to the tank on the map is not the same as having a safe path to the tank.
These changes have different verification states. Squad movement and combat improvements reached the main population on 28 October after repeated tests. Further healer protection, safer regrouping and responses to loose enemies were still being compared in the test groups at the checkpoint used for this article. A promising change does not become a reliable live feature because it has been written.
A success rate can hide an evening
One tested version recorded about 43% completion across 75 fair runs; a newer one recorded about 48% across 23. Different samples, different sizes. That is encouraging evidence to investigate, not a controlled proof of a five-point improvement.
And the harder problem was still waiting. At the morning checkpoint, the FB39 group had no clears in 52 fair attempts over two days. In 15 boss fights, Farren had not fallen below 62% health. The tank often died in about 20 seconds, usually after the Cleric.
Publishing that is less flattering than publishing only a boss kill. It is also much more useful if you are deciding whether to spend an evening testing a dungeon.
The current work keeps that squad near its agreed level range. Equipment improvements are allowed within the test limits, while survival, extra enemies and the approach to the fight remain under investigation. The result still has to happen in the dungeon.
What “the AI learned” means in this story
Nobody typed a dungeon transcript into a training job and received a wiser Cleric overnight. We examined failures, changed behaviour, repeated the fights and checked what actually happened. That is a development loop.
A model might phrase a resident’s reaction afterwards. It does not make an untested movement change safe, bring a dead healer back by sounding convincing, or turn an unfinished run into a clear.
What I want to earn is much smaller and more valuable: you call for help, the party arrives, and when an extra monster heads for the Cleric, somebody notices.
Based on dungeon test records and development checkpoints from 27–28 October. Counts describe those tests, not a promise about your next run.
Continue with what a resident actually remembers or the question about training on player data.