Script scope and portable paths
For offline data inspection, start with python load_tasks.py. It makes no model calls and does not generate or modify images.
GUI helper scripts now locate their data/ directory relative to the script, regardless of the working directory. Set GUI_ROOT to an absolute directory only if using a different local layout. migrate_chrome_data.py additionally requires MMA_CHROME_SOURCE_ROOT, pointing to an external raw Chrome export that is not bundled. No fictitious source directory is supplied.
The scripts/ directory includes historical generation tools and alternative evaluation scripts, not a single verified end-to-end release runner. Some generation tools require intermediate inputs such as 02_raw_popups/ that are not included. Do not run them merely to inspect the released tasks: they can call paid APIs or overwrite generated assets. The final task index points to already generated inputs.
Only run_frontend_benchmark.py is the maintained GUI runner with the fixed 209-task denominator. Other experimental variants, including run_frontend_benchmark1.py, retain their historical evaluation logic; the path cleanup does not migrate their metrics or validate their historical results. The main runner evaluates model-output coordinates; it does not physically execute clicks.
Configure provider credentials through the environment variables used by each script. No real key is bundled. Provider endpoints and model IDs in historical helpers are not a promise of present API availability. Run separate evaluations into separate output directories, and do not overwrite historical results when checking reconstructed inputs.
Path portability and Python syntax have been checked without importing or running the generation scripts. API behavior, external source exports and output equivalence are separate checks.