Fusion Kingdoms
Fusion Kingdoms is for people who play strategy games against the computer, like Age of Empires or RimWorld. In these games the "AI" is a hand-written script: every unit acts the same, and the enemy follows a fixed plan that players learn and beat. Real learned AI does exist, like AlphaStar in StarCraft or Sophy in Gran Turismo, but it's a black box you play against. You can't see it, shape it or own it. So players never get to work with real AI, only against scripts or sealed models.
Preliminary score
- Value for the user17.0 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.5 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing17.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility10.0 / 15
Data requirements, running costs and a clear next step.
- Originality8.5 / 10
A meaningful difference from common solutions.
In the final, the live jury awarded this project 3rd place with a score of 86.0 / 100.
What the judges wrote
- GPT-6 AstraPreliminary81/100
- Strengths
- Your central mechanic makes AI part of play: players fuse trained networks, mutate them and improve them through drills, rather than merely facing an opaque opponent. Your testing stands out for specific before-and-after results and candid failures, including misclassified units, ineffective healers and the weaker multitask model. Your compact networks, offline builds and headless simulation provide a credible technical foundation.
- To improve
- Your next playtest should measure enjoyment, strategic variety and player understanding against a fixed-behaviour baseline. Make your testing reproducible with saved seeds, per-scenario results and ablation tables, and expand the eight-match ruler comparison. Add an end-to-end recording showing a fusion or drill changing both measured skills and subsequent match behaviour. Quantify training time and hardware requirements, benchmark large-army runtime performance, and prioritise a native Windows smoke test.
- Claude Opus 5.5Preliminary76/100
- Strengths
- Your core idea is clearly executed: every unit runs its own small PPO-trained network (47→64→64→18). Weight fusion, mutation and in-game drills are actual gameplay mechanics, so the AI is the game rather than decoration. The live brain visualiser, which shows senses, hidden layers and chosen actions next to measured skill scores, is strong evidence that the networks drive behaviour. Your testing is unusually rigorous for a hackathon: a headless harness with 179 checks, 32-episode skill measurement, and ablations that took the ruler from 3/6 to 8/8 against the old AI. You also reported honest negatives, including the Thief+Guard fusion clash, the multi-task brain that lost 37 vs 47, and untested Windows.
- To improve
- Your user value rests on the claim that scripted AI is boring, but you show no evidence from actual players. Even five playtesters rating how predictable and learnable the AI feels against a scripted baseline would make the outcome measurable. Compare your rulers directly with a conventional scripted or behaviour-tree opponent, not only with your own earlier version, so the 'better than today' claim is grounded. State running costs and requirements explicitly: on-device inference cost per unit, training compute for 125M decisions, and build size. Then add a concrete next step beyond the multi-task brain fix, such as a playtest round or a Steam demo milestone.
The team
Fuad AliyevLead
product
Muhammedali Aliyev
developer
Nazrin Gasimova
developer