AI GamingFinalistGPT & Claude

Level-Forge

Level-Forge is for level designers at small puzzle-game studios and solo developers. These are the people who have to ship 50 to 200 hand-made levels for a game like Sokoban or Baba Is You, usually with no tools programmer on the team. Today they draw every level by hand and then play it by hand to find out three things: can it be won, is there a shortcut that skips the puzzle, and how hard is it really. Mistakes they miss are found by players after release. Asking a chatbot for a level does not help, because it does not know this game's real rules or file format, and nothing checks whether the level can be won. With Level-Forge the designer types a request in plain words, for example "a hard level about pushing rocks and changing the rules". The tool reads the game's own code to learn its real rules, builds a level in the game's own file format, and then wins it in the original game before handing it over. Each level comes with its solution, proof that it needs the skills that were asked for, and a measured difficulty. Requests that are impossible are refused, with the reason. We show it working on two different games. This concept can be expanded into almost all general puzzle genre games (including 3D).

Open the demo Video GitHub

Preliminary score

83.5/ 100Strongaverage of 2 judges

Value for the user20.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI23.5 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing18.0 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility13.0 / 15
Data requirements, running costs and a clear next step.
Originality8.5 / 10
A meaningful difference from common solutions.

A finalist: the live jury scored it 61.0 / 100 on stage.

What the judges wrote

GPT-6 AstraPreliminary84/100
Strengths
Your engine-replayed solutions and skill-ablation checks address a concrete designer problem: creating solvable levels that cannot bypass the requested skills. Your 200-request comparison, 100,000-move differential test and specific failure-and-repair examples provide unusually strong testing evidence, with full validation improving from 59% to 95.5%. Your second-game adapter and explicit distinction between solvable and fully validated outputs strengthen the prototype within its stated scope.
To improve
Your next evaluation should run the full Gemini path on held-out, human-written requests, reporting intent accuracy, validation rate, latency and cost separately from the keyword baseline. Test with designers and players to measure authoring time, level variety and agreement with your difficulty score. Show a fresh-request, end-to-end generation recording alongside the hosted demo's recorded runs. Before claiming broad puzzle-game coverage, expand beyond the current Baba pattern family and document the human intervention needed to onboard another game.
Claude Opus 5.5Preliminary83/100
Strengths
Your verification loop is the heart of the project, and you tested it seriously. Every level is replayed in the original engine, ablation search proves each requested skill is necessary, and your fast solver was differential-tested against the real engine on 100,000 moves. Your benchmark has a clear structure: random tiles vs. builder without checking vs. full pipeline, reaching 95.5% fully validated on Baba Is You and 96.7% on Sokoban. Your honesty also stands out: you documented the 'pushing is required' shortcut, the long-but-easy 'hard' levels, and the Gemini script that silently saved keyword results. This makes your numbers far more credible than typical success claims.
To improve
At runtime the AI is mostly classical search, and the language model is optional and unbenchmarked. Benchmark Gemini (or a faster model) against the keyword parser on paraphrased requests to show what it adds. The rule book was written by a coding agent working with you, so demonstrating an unattended blueprint for a third game would back up your generality claim. Your submission text says the approach extends to 3D games, while your deck says 3D is untested; keep these consistent. Calibrating your difficulty score against real players and timing a designer's manual workflow would turn the 'compared with today' table into measured outcomes.

The team

Chingiz AslanovLead
data
1) Nural Shukurlu
data
Alfaddin Hamidov
developer

Disclosure

Required by the rules. Judges verify it against the repository.

Models: Anthropic Claude (coding assistant: read the game code and wrote the game blueprint format). Google Gemini, free key (reads requests and names levels; works but slow, so the demo and all test numbers use the test version without it).
Data: No datasets. Only the code and built-in levels of two open-source games (MIT licence): utilForever/baba-is-auto and xbandrade/sokoban-solver-generator. Test requests were generated by our own script.
Components: Python 3, C++ (g++), pybind11, pygame, pygame-widgets, numpy, GitHub Pages. No templates.
Built during the event: Confirmed

Quality testing

WHAT WE TESTED 1. 200 requests on the Baba Is You engine, made three ways and judged by the same checker: - random tiles: 0% winnable - our builder with checking switched off: 92% winnable, 59% passed every check - Level-Forge: 100% winnable, 95.5% passed every check "Passed every check" means the level can be won, is as hard as requested, and cannot be won without the skills that were asked for. 2. Can our checker be wrong? Every level we marked as winnable was replayed a second time through a separate route into the original game: 0 of 200 failed. Our fast solver and the real game were compared on 100,000 random moves and never disagreed. On 120 levels we broke on purpose, they gave the same verdict in all 117 cases where the real game could finish its search. 3. The rule book: 30 rules read from the game's code, each one tested in the real game. 60 of 60 automatic tests pass. 4. Impossible requests: 19 of 19 were refused with a reason, and 0 of 10 valid requests were refused by mistake. We wrote these requests ourselves while building the checker, so this shows it keeps working, not how accurate it is on new requests. 5. A second game (Sokoban) through the same unchanged program: 60 requests, 96.7% winnable and passing every check, 0 of 58 wrongly marked as winnable. WHAT BROKE - The real game engine was too slow to search (about 2,000 positions per second). We wrote a fast copy of its rules and still replay every solution in the real game. - We claimed "pushing is required" for some levels. Our own search proved us wrong: a rule word could be sunk instead of a rock. We changed how levels are built. - Our first "hard" levels were long but easy. Difficulty is now measured by how many traps a level has, not by its length. - The second game passed only 68% at first, because boxes were placed in corners they could never leave. Fixed, now 96.7%. - About 1 in 10 hard requests still misses a check. Those levels are delivered with a lower label and the reason. Not fixed. - Our first Gemini run failed on a wrong key, and our script quietly saved the keyword reader's results as if they were Gemini's. We caught it, deleted those numbers, and the script now refuses to do that. Gemini works now, but at 7 to 10 seconds per call on a free key it was too slow for the 200-request test, so it is not benchmarked. COMPARED WITH TODAY By hand, a designer learns whether a level can be won, or has a shortcut, only by playing it. A chatbot can write something that looks like a level but does not know this game's rules, and nothing checks the result. Level-Forge returns a checked level in under a second for most requests (about 7 seconds for hard ones in our test), with the solution included. WHAT WE DID NOT SHOW We did not time human designers, so we do not claim hours saved. Our difficulty score has not been compared with real players. The test requests came from our own templates, so the numbers show reliability on the kinds of request the tool knows. We tested two games, both 2D and turn-based.

Worth 20 of the 100 points.

Running it

The demo link needs no setup or login. Pick a request on the left, press "Play solution" to watch the computer win the level, or press "Play it yourself" and use the arrow keys. In "Check your own level" you can edit a level and the solver runs live in your browser. To run the full app on Linux or macOS (about 2 minutes): git clone https://github.com/Maced78/level-forge.git cd level-forge bash build_engine.sh bash run_demo.sh Then open http://localhost:8765. It needs g++, python3 and python3-dev. On Windows, run the same commands inside WSL (Ubuntu). No credentials are needed. Without an API key the app runs its test version, where a built-in keyword reader understands the request instead of the language model, and it says so at the top of the page. The demo uses this version on purpose, because Gemini takes 7 to 10 seconds per call on a free key. To switch Gemini on, put GEMINI_API_KEY=... in a file named .env.

Project ownership

The team keeps the rights to this project. Judges were given access for verification only.