The Spire bot now beats Act 3 in about half its games
Ten days on: a bot that looks ahead between fights, a much bigger fight network, two bugs found in real games, and what still ends its runs.
Written on November 7 from the games played up to that afternoon, and published on November 19. What came next, the fourth act and the Heart, is in the next post.
Two weeks ago our Slay the Spire bot beat Act 3 in about a third of its games. It now beats it in about half. As before, every game here is a real one: Ironclad, Ascension 20, played in the real game through a mod, and you can replay any of them. There are 1,640 of them now.
Where it stands
Since November 3 the bot has won 220 of 448 finished games (49%). The bots playing since November 6 have won 38 of 69 (55%). In the first post’s last version it was 37%.
Share of finished A20 games that beat Act 3, by the day the session started; the lines are 95% intervals. A day is 75–150 games, so a single day moves by ±10 points on luck alone. November 7 is a part day (36 games).
Data table
| Day | Act 3 beaten | lo | hi | runs |
|---|---|---|---|---|
| 2026-10-23 | 2.6% | 0.5% | 13.5% | 38 |
| 2026-10-24 | 17.0% | 11.1% | 25.0% | 112 |
| 2026-10-25 | 13.3% | 6.9% | 24.2% | 60 |
| 2026-10-26 | 22.1% | 15.2% | 31.0% | 104 |
| 2026-10-27 | 28.8% | 21.9% | 36.8% | 139 |
| 2026-10-28 | 31.3% | 24.5% | 39.1% | 150 |
| 2026-10-29 | 35.9% | 28.1% | 44.5% | 128 |
| 2026-10-30 | 37.3% | 27.3% | 48.6% | 75 |
| 2026-10-31 | 32.2% | 24.6% | 41.0% | 121 |
| 2026-11-01 | 49.4% | 38.6% | 60.2% | 79 |
| 2026-11-02 | 44.6% | 36.1% | 53.5% | 121 |
| 2026-11-03 | 51.6% | 41.5% | 61.6% | 91 |
| 2026-11-04 | 45.5% | 37.3% | 54.0% | 134 |
| 2026-11-05 | 45.2% | 36.0% | 54.8% | 104 |
| 2026-11-06 | 49.0% | 39.3% | 58.7% | 98 |
| 2026-11-07 | 66.7% | 50.3% | 79.8% | 36 |
These games still stop at the end of Act 3: the fourth act and the Heart stay locked for now.
What changed
The bot is two parts: one plays fights card by card, the other makes every choice between fights (paths, card rewards, campfires, shops, events). Both learn in a simulator of the game and then play real games. Since the last post:
- Oct 30 and Nov 2: more practice on itself. Two more rounds of training the choice model on its own simulated games, now with every elite fight played out instead of estimated (the lesson from the first post).
- Nov 3: it looks ahead between fights. Before a choice, it plays the next floors out in the simulator, several times over, instead of trusting a single estimate of each option.
- Nov 4: a retrained fight judgement. Two new versions of the network that judges how a fight is going. In real games this version did worse than the one before it (41% against 55%), though the samples are small and the intervals overlap.
- Nov 5: it remembers the act. Its lookahead used to start every act afresh, as if no fights had happened yet. It now knows which fights and elites the act has already thrown at it.
- Nov 6: a much bigger fight network, running on a graphics card. In 600 test runs in the simulator on fresh seeds, it beat Act 3 in 357, against 313 for the network before it. On November 7 its fight search was spread over several threads: about 8× faster, same moves.
Finished real games of each version since the first post, with 95% intervals. Versions that played fewer games have wider intervals; the data table has the counts. The full list of versions is on the Spire page.
Data table
| Bot version, from the day it started playing | Act 3 beaten | lo | hi | runs |
|---|---|---|---|---|
| v6 Oct 29 | 37.4% | 30.5% | 44.9% | 171 |
| v7 Oct 30 | 40.6% | 35.0% | 46.5% | 278 |
| v8 Nov 2 | 40.5% | 30.1% | 51.9% | 74 |
| v9 Nov 3 | 54.8% | 44.7% | 64.6% | 93 |
| v10 Nov 4 | 40.9% | 33.0% | 49.2% | 137 |
| v11 Nov 5 | 50.3% | 42.4% | 58.3% | 149 |
| v12 Nov 6 | 51.6% | 39.6% | 63.4% | 64 |
Two bugs the real games showed
Skipped chests. The bot remembered that it had opened the chest on a floor by the floor’s number alone, and that memory carried over from one run to the next. So now and then it walked past the Act 1 treasure chest (always on floor 9) as if it had already opened it: 185 of 1,560 floor-9 chests in all real games so far were never opened, and a few shops were never entered. Fixed on November 7.
Seeing the future in practice. In the simulator, the bot’s lookahead could foresee the random outcome of the room it was standing in: what Neow’s bonus or an event would really give. That made its simulated practice runs easier than real ones: they reached the bosses with more cards and relics than real runs at the same HP. And in real games, its lookahead planned as if it knew how the room would turn out. It now redraws those outcomes every time. Also fixed on November 7, so it’s too early to see the effect in real games.
What still ends its runs
The fight that ended each lost real game since November 3 (228 losses), by act: the act's boss, or any other fight (hallway fights and elites).
Data table
| kind | Runs lost | |
|---|---|---|
| Act 1 | Boss | 33 |
| Act 1 | Other fights | 10 |
| Act 2 | Boss | 53 |
| Act 2 | Other fights | 57 |
| Act 3 | Boss | 56 |
| Act 3 | Other fights | 19 |
Bosses end 62% of its lost runs. In Act 3 it’s three in four: the Awakened One, Time Eater and Donu and Deca. Act 2 is split: the Collector and the Bronze Automaton on one side, and elites like the Book of Stabbing and the Slavers on the other.
One choice stands out. At the campfire just before a boss, the bot can rest (heal 30% of its max HP) or upgrade a card. With half its HP or so, it often upgrades. In the real games since November 3, at 40–60% HP:
Real games since November 3, with 95% intervals. The bot upgraded 103 times and rested 88 times.
Data table
| At the campfire before a boss, with 40–60% HP | Beat the boss | lo | hi | times |
|---|---|---|---|---|
| Rested | 90.9% | 83.1% | 95.3% | 88 |
| Upgraded a card | 78.6% | 69.8% | 85.5% | 103 |
Real games alone can’t settle this: the bot chose, and it may upgrade when its deck looks strong. (Below 40% HP the raw numbers even point the other way, for that reason.) So we took 831 real campfires before a boss and played the boss out from each one in the simulator, both ways. Resting at 40–60% HP beat the boss about 12 points more often there too, but the bot’s choice model counted the difference as small. A version that plays the boss out before choosing at these campfires is being tested now.
Next
Test the campfire fix, see what the two November 7 fixes do in real games, and then unlock the fourth act and the Heart. The bot keeps playing in the meantime: the Spire page and the run browser update with every refresh of the data.
Since then
Added November 19. The two November 7 fixes did not move the win rate on their own: the versions from November 7 to 12 beat Act 3 in about half their games, like the ones before. What moved it was a different kind of bot, one that taught itself the game in the simulator, and with it the fourth act. That story is the next post.