Jack Tricker / Writing

Spire Network 5/5: How strong is it, and what comes next

Jack Tricker, 17 November 2026


Part 5 of 5. Part 4 covered rules and a pro's runs.

Play it

The engine compiles to WebAssembly, so the whole thing runs in a browser tab at spire-viewer.pages.dev. The engine is 1.8 MB, a network's weights are 4.5 MB, and a decision takes 5.6 ms. You can watch the networks' best runs, play any character on any seed at any ascension, and ask the network what it would do and how likely it thinks you are to win. You can also go back to any decision and choose differently, and the run continues from there as a variation, like analysing a game on a chess site.

Where it stands

When I posted the project on Reddit, one of the criticisms was that the models were weak. So I looked up what strong is. These are win rates against the Heart at A20, as I found them in November 2026.

PlayerA20 Heart wins
The best tracked human, by character (a community tracker)79 to 96%
Baalorlord, lifetime, by character (the same tracker)53 to 76%
destire-mio's search bot, Ironclad, 501 runs in the real game50.1%
The best language-model agent in Epoch AI's tests, by character0 to 33%
Everyone who played A20 in Mega Crit's 2020 data, by character1.9 to 7.9%
My Watcher network8.0%
My Defect network0.9%
My Silent network, with the look-ahead and a key rule1.7%

The Ironclad isn't in the table because it was never trained for the Heart. Its 49.5% from part 2 is for beating the Act 3 boss, which is an easier target than the search bot's 50.1%.

I haven't found a published network that plays A20 with no search and does better than these. Against what players mean by strong, the criticism was fair. Every strong bot searches in its fights, and so did mine when it reached 49.5%.

What limits it

More of the same training has stopped helping. Over the Watcher's 944 million training decisions, the network won 73.6% of fights from human runs in the first half and 73.8% in the second. The Silent's figures over 942 million decisions were 76.7% and 76.9%.

Bigger networks didn't help under this training either. From scratch, a network of 0.9 million parameters beat ones of 1.3, 2.9 and 4.5 million.

Search helps a lot. The same Ironclad weights go from 27.1% to 49.5% with a two-turn look-ahead. And the fights it loses can mostly be won: when it replayed lost fights from human runs eight times with different draws, it won at least one replay of 74% of them.

The decks are narrow. Perfected Strike is in 52% of the Ironclad's Act 3 decks and 6.5% of human winners'. The Silent's fighter beats the Heart 4.3% of the time with the decks its own deck network builds and 22.4% of the time with human winners' decks.

What comes next

The next experiment on my list is teaching a search to a network properly. My attempts so far used at most 438,000 positions, which is small. The plan is to label a much larger set of positions with the look-ahead's values, train networks from 0.9 to 12 million parameters on them, and see whether any of them learns to play like the search with no search running. I estimated $35 to $125 of rented graphics cards for it. It's on hold while I decide.

How it was built

The repository had 1,297 commits in its first month, and I didn't type most of them. Claude Code wrote the code, ran the experiments and wrote the reports. I chose what to build and when to stop.

A few habits kept that honest:

One bug hid for a week. Once the deck network and the fighter were separate files, the search read its reward scale from the wrong one (1.0 where it should have been 0.32), overvalued potions, and won 0 to 6% of runs where it should have won about 50%. My old evaluation script always passed a flag that happened to hide it, and a new command without the flag exposed it. I now rerun the exact old command before I blame new code.

The code is at github.com/jackybabes/spire-engine under the MIT licence, with a model card for each network.

[ Return to top ] [ All writing ]


Jack Tricker / me@jacktricker.com
Last modified: 17 November 2026