The benchmark asks agents to spend a fixed budget well
Tokenbrawl places two agents in a deterministic one-on-one fighting game. At each decision point, both agents receive an observation and return one legal action. Each response spends from a finite Token Bank. If an agent runs out, it enters Reflex Mode, where its next decision is capped at eight tokens.
This turns inference budget into part of the game state. A model can spend more on a complex decision, but doing so leaves less capacity for later turns. The benchmark is designed around a strategic tradeoff rather than a single prompt-response score.
Latency is separated from the outcome
The match runner waits for both agents at each decision point before advancing the simulation. The simulation advances in ticks rather than wall-clock time, and the replay player advances by frame count. The command log does not need response latency to determine the result.
These controls reduce one source of unfairness: a faster deployment should not win solely because it answered sooner. The deterministic environment and replay log make it possible to inspect what happened and reproduce the match under the same configuration.
The demo is not evidence of a completed tournament
The repository is explicit that its showcased animation is a real engine replay between a scripted stand-in and a baseline bot. It is not a live model match, and a tournament has not been run yet. That boundary is important because an attractive replay can otherwise look like a benchmark result.
The harness is most valuable when an evaluation includes enough controls to interpret the result: a fixed environment, legal action constraints, deterministic replay, model and prompt versions, and a clear account of what was actually run. Until a model tournament is executed and reported, Tokenbrawl is an evaluation system in development, not a model leaderboard.
INSPECT THE IMPLEMENTATION
Read the code and project notes
The details above describe the public repository and project status as reviewed for this portfolio. Follow the source to inspect the implementation, setup instructions and current limitations.
Tokenbrawl harness and replay details
