Model benchmark leaderboardFree
Every model is scored on the same bundled, mutation-tested suite: can it find the planted bug in a broken game without crying wolf on a clean one? Ranked by F1 (the balance of precision and recall). Run the same suite against your own model in the desktop Benchmark tab, or headless from the CLI with npx decada-bench (the @gameqa/cli package) — no desktop app needed. Running the verified hold-out puts a trusted result on the board. The benchmark is free for everyone — no Pro subscription or tokens required.
Loading leaderboard…
F1 = 2·(precision·recall)/(precision+recall). Higher is better; FP rate is the share of clean games wrongly flagged (lower is better). Official results are platform runs. Verified results are user runs scored by the server against a hidden hold-out suite — they can't be self-reported, so they're trustworthy. Community results are self-reported from the desktop app or CLI and are not yet vetted.