How detection is tested

The method

1. 1

   Real recordings

   The test set is real gameplay footage across five game generations, from Fire Red/Leaf Green (2D) to Sword/Shield and Scarlet/Violet (3D): soft resets, starters, fishing, breeding and fleeing wild encounters.

2. 2

   Templates like a user makes them

   For every recording, detection templates are created exactly the way a user would create them in the app, no hand-tuned special settings.

3. 3

   Hard pass criteria

   Every frame of every clip is scored. The suite only passes if every real encounter is found, the full video scan counts the exact number of encounters, and nothing is ever counted twice. The encounter simulation follows the app's real adaptive polling.

The test recordings were kindly provided by LEOsMIND. Huge thanks, this footage has substantially improved Encounty's development. LEOsMIND on YouTube.

Current results

Last run: …

Detection accuracy is hardware independent: the same scoring math returns the same result on any machine. This suite also runs entirely without a GPU, using the CPU pipeline.

Signal vs. noise per scenario

For every scenario, the weakest real encounter frame must still score clearly above the strongest non-encounter frame. The distance between the two bars is the safety margin.

Same result on CPU and GPU

The WebGPU shader pipeline is checked against the CPU pipeline on the same frames from the same recordings. Both must agree, so it does not matter which one your machine ends up using.

What gets tested

The test recordings contain game footage and therefore cannot be published; the method and the latest results are documented here instead, and the suite runs in CI against the privately stored fixtures. Questions or want to dig deeper? Open an issue on GitHub.