The numbers
Six full arena clears by the designer, thirteen by the bot, every event logged on both sides. The verdict was not close: the human clears the thirty-three-monster gauntlet in about ninety seconds taking ~175 damage; the bot needs twice the time and twice the damage.
The interesting part is why. It isn't reflexes — the bot dodges more, rolls more, and reads telegraphs frame-perfectly. It's tempo. Both players land almost exactly the same number of basic attacks per arena; the human just does it in half the time, avoiding melee by spacing rather than i-frames (six arenas, not one single i-frame escape), and calmly eating a few arrows as the cost of never stopping. The bot, meanwhile, buys its safety with long retreats that spend half the clock walking instead of fighting.
The bug only a comparison could catch
One bot batch posted suspiciously beautiful numbers — and the diff showed why: across five runs, not one enemy arrow connected. The crossbow bolt's new arrow model carried an invisible leftover collider, and the projectile's own hit-check was colliding with it and quietly deleting itself at launch. Rangers had been firing blanks all batch. No human eye would spot that at triple speed; two datasets side by side made it obvious in seconds. Fixed, and the analyzer now flags any run where bolts stop connecting.
The student copies the master
The designer watched his own play and named the real difference: he circles packs to bunch them up — keeping his distance while herding monsters into a clump where one big cleave hits five skulls — where the bot ran straight away and stretched them into a line of single targets.
That observation became bot doctrine the same evening. The new herd-orbit behavior circles the pack's center of mass, holding the nearest skeleton right at sword's reach, letting the monsters' own chasing compact them. First batch with it: five full clears out of five — the bot's first perfect batch ever — with a new speed record. The orbit logic is still rough (it stutters in and out of the mode), which is next on the list, and that list now has a mission statement: keep iterating, batch after batch, until the machine plays this arena better than its teacher.