Jack Tricker, 14 November 2026
Part 4 of 5. Part 3 covered the other characters.
Every number here is an A20 Ironclad run that beats the Act 3 boss, measured on 1,024 seeds the networks never trained on and compared seed by seed. The yardstick is the pair from part 2 with no look-ahead: the deck network chooses between fights, the fighter plays the fights, and together they win 26.1%.
A network can't tell you why it skipped a card, and a rule book can. I wanted readable rules for every choice outside fights (card rewards, shops, the map, rest sites, events, boss relics, Neow), with the network kept only for the fights. Rules would also be small enough to run in the browser and could explain each choice as they made it.
| Rules | Runs won |
|---|---|
| Written from human tier lists and human winners' pick rates | 0.6% |
| Counted from the network's own choices in 8,192 recorded runs | 5.6% |
| Fitted to those choices, the way chess ratings are fitted to results | 15 to 16% |
| Fitted, with the network still picking the cards | 19 to 20% |
| The deck network | 26.1% |
The human rules lost 69% of their runs in Act 1. What a strong human picks assumes that a strong human plays the fights that follow.
Fitting rules to the network's choices gives a score to chase: how often the rule picks what the network picked. That score misled me more than once. A better card rule raised agreement from 74.7% to 78.7% and wins from 16.7% to 17.4%. A refit of the shop rule agreed with the network more and won fewer runs, 21.7% against 24.0%. That one turned out to be a bug: my recorder had been saving the shop's stock over the list of potions the player held.
The quickest way to find bugs like that was to play the rules and the network on the same seed and read the two runs side by side. Ten minutes of that showed the rules never drank Fruit Juice. The code was matching "FruitJuice" and the game calls it "Fruit Juice".
With the fitted rules on one screen and the network on the rest, each screen moved the win rate by this many points:
| Screen | Change |
|---|---|
| Neow | +0.1 |
| Map | -0.2 |
| Events | -0.5 |
| Boss relics | -1.0 |
| Upgrades and removals | -1.5 |
| Rest sites | -1.8 |
| Shops | -2.1 |
| Card rewards | -6.3 |
Apart from card rewards, each screen is too close to the network for a 1,024-run test to tell them apart. With all of them on together the rules lose 9.9 points. Part of that is small losses adding up. Part is that the screens depend on each other: the fitted map rule sends decks built by the rules into early elites, and Act 1 deaths went from 115 to 302. And the fighter learned to fight with the decks this deck network builds, so decks built another way are less familiar to it.
I added dials to the rules for how often they rest, fight elites and leave shops. None of them helped by more than the noise, and the best combination reached 17.6%. The loss comes from which card, which elite and which purchase, and the readable weights can't see enough of the run to choose.
To put a price on card picks I had the network's pick swapped for a random other option some of the time. Swapping 10% of picks cost 4.9 points, 20% cost 8.8 and 30% cost 14.0. A run opens about 18 card rewards, so each swapped pick costs 2.5 to 3 points of win rate.
The fitted card rule disagrees with the network about one time in five and loses 6.3 points, which is what random swaps 14% of the time would cost. Its disagreements are only about a third cheaper than random picks. Adding 11,000 weights for pairs of cards that go together raised agreement to just 80.6%.
Baalorlord is a streamer who posts every run on baalorlord.tv with the game's own run file. I downloaded 3,932 of them at one request a second. Of those, 808 are A20 Ironclad runs, and he beat the Heart in 60% of them. The files hold his choices between fights and nothing from inside the fights.
His boss fights were no easier for my fighter than other human winners' fights. It won 60.4% of his and 61.6% of theirs, and with a four-turn look-ahead 74.4% and 74.2%. So his 60% comes from how he plays his decks, since the decks themselves are no easier to win with.
A network trained to copy his choices went from agreeing with him 53.9% of the time to 74.1%, and it won fewer runs: 22.1% against 26.1%. Act 1 deaths rose from 115 to 202. His early choices need a player who can win hard fights with them.
So I moved the fighter towards him. It trained for 500 more updates in whole runs where the network that imitates him made the choices.
| Choices made by | Old fighter | Adapted fighter |
|---|---|---|
| The network imitating him | 22.1% | 24.9% |
| The deck network | 26.1% | 26.8% |
The gain with his choices holds up seed by seed: 113 seeds were won only by the adapted fighter and 84 only by the old one (p 0.046). With the deck network's own choices nothing changed. His choices still trail the network's, 24.9% against 26.8%, and the deaths fall in different places. Runs on his choices die more often in Act 1 (210 against 146) and less often in Act 3 (239 against 301).
A fighter also has to learn inside whole runs. The ones I trained only on practice fights won 11.8% of runs, because they drank 3.6 potions a run in hallway fights where the run-trained fighter drinks 0.6.
Next: part 5, how strong all this is, and what comes next.
[ Return to top ] [ All writing ]