AI GamingFinalistGPT & Claude

ProofPlay

ProofPlay is for PC racing players who want to improve their lap times but cannot identify which braking or throttle decisions are costing them time. Existing statistics show where they were slow, while coaching advice often recommends changes without checking whether they actually help. ProofPlay replays the recorded race, tests alternative inputs in the game’s own physics, and shows only changes that saved time without crashing. Players receive specific corner-level fixes and a next-race plan whose combined effect has also been tested.

Video GitHub

Preliminary score

84.5/ 100Strongaverage of 2 judges

Value for the user21.5 / 25
A specific problem and a clear outcome from solving it.
Prototype and use of AI24.0 / 30
A working core scenario, and what AI actually contributes to the result.
Quality testing18.5 / 20
Tests, examples of failures, and comparison with the current approach.
Feasibility11.5 / 15
Data requirements, running costs and a clear next step.
Originality9.0 / 10
A meaningful difference from common solutions.

A finalist: the live jury scored it 66.0 / 100 on stage.

What the judges wrote

Claude Opus 5.5Preliminary83/100
Strengths
Your core idea of showing only advice that has been re-simulated in the game's own physics is distinctive and well executed. It is backed by deterministic replay with end-state hashing, 11-variant what-if search per corner, and robustness windows. Your benchmark against a rule-based coach and GPT-5 mini on 14 races (31/31 vs 76.5% vs 37.5%) is a real comparison with how coaching is done today. You honestly frame the 100% result as a product of filtering rather than a guarantee. Documenting failures, such as combined fixes making a lap 0.69 s slower or crashing, and the open-loop drift that led you to one-lap plans, is exactly the kind of evidence that builds trust.
To improve
Your benchmark relies mostly on bots with planted mistakes. Running even a small study where human players follow the fixes, and measuring their real lap-time change, would show whether proven-in-simulation translates to proven-for-humans. The AI contribution is search and optimisation over physics rather than learned models, so explain clearly why this counts as intelligence and where it beats a brute-force sweep. Feasibility is tied to your own deterministic game. Spell out a realistic path to other titles (for example, games with deterministic replays or telemetry APIs) and who would pay for it.
GPT-6 AstraPreliminary86/100
Strengths
Your physics-based search produces actionable corner-level changes, while robustness windows and combined-plan replay address real ways individual tips can fail. Your three-coach benchmark, ten tamper cases and explicit slower/crashing combinations provide unusually specific testing evidence for a hackathon submission. Your API-free, headless analysis also has concrete latency and memory measurements.
To improve
Your next benchmark should use held-out human races, report how often no advice is returned, and measure actual next-race gains rather than only gains in the filtering simulator. Your leaderboard claim should distinguish replay consistency from authentic play: test an attacker who regenerates both the session and its hash, rather than only editing individual fields. Your deployment plan should name a target CPU, a per-race compute budget and a concrete human-pilot milestone.

The team

Omar Musazade
product
Sanan Hajiyev
developer
Rufat DostaliyevLead
data
Farrukh Mammadli
data
Ravan Khanbabayev
business

Disclosure

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

Models: The core product uses deterministic physics-based counterfactual search and a six-attribute skill profile calibrated against bot runs. It does not use an LLM for player-facing recommendations; explanations are generated from measured results using deterministic templates. Azure OpenAI GPT-5 mini was used only as a comparison coach in the evaluation benchmark.
Data: First-party race recordings from RIVAL LINE, including per-step player inputs and race configuration; synthetic bot races with deliberately planted braking mistakes; clean Normal and Hard bot runs; and a small number of human-driven races. The benchmark contains 14 races, including 12 with planted mistakes. Track geometry, vehicle specifications and tuning come from our own game content. No screen recordings or external training dataset are required.
Components: C++20; a fork of Wicked Engine 0.72; Jolt Physics; miniaudio; nlohmann/json; Python standard library; SQLite; HTML, CSS, JavaScript and Canvas 2D; Phosphor Icons; Google Fonts; Blender-generated assets; CMake and Visual Studio 2022. Development also used the Claude Code Game Studios workflow. Pre-existing work: RIVAL LINE, our original combat racing game, existed before the hackathon. ProofPlay adds recording, replay verification, race analysis, counterfactual coaching and the dashboard on top of that game.
Built during the event: Confirmed

Quality testing

We evaluated three coaches on the same 14 races, including 12 with deliberately planted braking mistakes, and re-simulated every recommendation in the game’s physics. ProofPlay’s 31 recommendations all saved time, averaging 1.054 seconds per tip. The rule-based coach improved 12/32 recommendations (37.5%), while GPT-5 mini improved 26/34 (76.5%). The rule-based coach produced 19 recommendations that worsened performance or crashed; the LLM coach produced seven. ProofPlay’s 100% result applies to this simulation benchmark and reflects its policy of filtering out unsuccessful candidates; it is not a guarantee of improvement for every human player. We also tested replay integrity. The genuine control session was accepted, while all 10 manipulated sessions were rejected, including altered times, positions, inputs, difficulty, lap count and vehicle parts. A major failure was assuming that individually successful fixes could simply be combined: one combination made a lap 0.69 seconds slower, and another crashed. We therefore added combined-plan re-simulation. Long open-loop replay could also drift or get stuck, so plans are limited to one lap. Compared with statistics-only coaching, ProofPlay checks proposed changes before presenting them. Remaining limitations include a small human sample, Windows-only analysis binaries, support for RIVAL LINE only, and counterfactual changes limited to braking and throttle.

Worth 20 of the 100 points.

Running it

Requires Windows 10/11 and Python 3.10+. Building from source also requires Visual Studio 2022 with the C++ workload and CMake 3.19+. Follow the README’s “Building from source” instructions, including cloning submodules and applying the four Wicked Engine patches. Build the racer_sandbox and proofplay targets in Release mode. Start the dashboard with python proofplay/dashboard/server.py, then open http://localhost:8787. Click “Start session,” enter a driver name, launch build/game/game/sandbox/Release/racer_sandbox.exe, and finish a race. Results appear automatically; full analysis typically takes 45–130 seconds and can take approximately five minutes for longer races. The core product requires no API keys or login. Azure OpenAI credentials are needed only to reproduce the optional LLM comparison benchmark. Currently, only RIVAL LINE is supported, and unfinished races are not saved.

Project ownership

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