Jack Tricker, 11 November 2026
Part 3 of 5. Part 2 covered the Ironclad.
From here a win means beating the Heart at the end of Act 4, which takes the three keys. Getting the networks to go for it was most of the work.
The first Silent network started from the Ironclad's weights and was trained the same way, with the Act 3 boss as the win. It won 76% of runs at A0 and 3% at A20, and it beat the Heart zero times at any ascension because it never took the Sapphire key. From where it stood that was correct. The keys cost a rest, a relic and a harder elite fight, and all they bought was a fourth act to die in.
The Defect was trained from scratch with 1 point for the Act 3 boss and 7 more for the Heart. That alone didn't make runs take the keys, so I added a bonus per key. At high ascensions, runs then grabbed keys and died before Act 4, because the bonus was paid whether or not the run got there. Paying it only to runs that reached Act 4 fixed that.
The Watcher needed the opposite. With the bonus paid only in Act 4, she held about one key a run and stayed there. Paying for every key held took her from 1.06 to 1.87 keys a run, and from 0.6% to 6.2% of runs reaching Act 4, within 125 updates. A larger Watcher network on that same bonus later went back to grabbing keys and dying. I never found one setting that worked for every network.
The released Defect beats the Heart in 27.5% of runs at A0, 10.4% at A10 and 0.9% at A20. The Watcher beats it in 43.8%, 26.3% and 8.0%. For scale, in Mega Crit's 2020 data 1.9% of everyone's A20 Defect runs and 7.9% of their A20 Watcher runs beat the Heart.
The Defect plays her orbs passively. She plays Dualcast only 27 to 51% of the times it's in her hand, and given a deck built around Dark orbs she fills her slots with Frost.
The Watcher handled stances well in fights and wouldn't build a stance deck. She took Mental Fortress on 5% of offers where human winners take it on 46%. One power card added to a thin attack deck doesn't pay for itself, so the network never took the first step. The fix was to start a fifth of training runs with Mental Fortress+ and Rushdown already in the deck, so that it played the engine and learned what it was worth.
The Silent's trouble is that poison pays over several turns. A look-ahead of one turn made her worse: 0.3% of A20 runs past the Act 3 boss against 2.7% without it. Two turns helped (3.6% against 1.7%) and three helped more (4.5%).
She also fell for Shivs. Blade Dance is a cheap common that turns up often, and she took it 87% of the time where human winners take it 4 to 8%. In the human data, players with a Shiv deck at the Act 2 boss went on to beat the Heart 2.9% of the time. Players with a Grand Finale deck did it 39.7% of the time.
In practice fights, a Smoke Bomb on turn 1 against three Darklings counted as a win at full HP, until I made escaping count as a loss.
One Defect run found an infinite combo and played cards at the Heart on 1 HP. The Heart caps the damage it takes each turn, so the turn went on until the run hit the limit of 4,000 decisions. Turns are capped at 100 moves now.
Forcing human winners' card picks on the networks made all three worse. The Watcher fell from 9.8% to 3.5%. Their fighting can't use the cards the humans would have taken.
From 5 to 9 November everything went on one goal: A20 Heart wins for the Silent.
A new, larger fighter trained from scratch beat the old one on fights from human runs, 75.9% against 69.7%, and then died earlier in whole runs, around floor 19 against floor 25. A practice fight ends when the fight ends, so it had learned to drink everything: 3.08 potions a run in hallway fights against the old network's 0.58.
These are the changes that moved the number:
It ended at 1.7% of fresh A20 runs beating the Heart, 51 of 3,072. The average for human Silent players at A20 is 3.3%. The network reaches both Act 3 bosses three times as often as those players do, then wins 17% of its Heart fights where they win 68%. I stopped the training on 9 November after the numbers had stopped moving.
The human data needed cleaning along the way. About 8.5% of the "winning" Silent runs look cheated, with 1,000 gold inside three floors or 30 relics.
Next: part 4, a rule book in place of the network, and a pro's runs.
[ Return to top ] [ All writing ]