Adapters and validation

Gazepoint

adapt_gazepoint_samples() converts declared Gazepoint-style columns to the canonical schema. It defaults to fractional screen coordinates and seconds because those are common Gazepoint export semantics, but every relevant source column and unit can be overridden explicitly.

The adapter requires screen_size_px when coordinates are normalized. It never guesses the screen resolution.

Existing Python ecosystem tables

adapt_processed_table() is the bridge for eyeprocesspy, gpbiometricspy, and other preprocessed tables. The caller supplies the participant, trial, time, x, and y columns explicitly. This avoids coupling GazeForge to unstable upstream private column conventions.

Leakage-safe validation

grouped_event_cross_validate() uses GroupKFold and fits a fresh event classifier in every fold. The default grouping unit is participant_id.

assert_no_group_leakage() can additionally be applied to participant, stimulus, session, or other grouping columns before any train/test evaluation.