task47_0809/scripts/README_followups.md

Follow-up experiment scripts (Task 4 deep dives + Task 7 prototype)

These accompany the original run_ijepa_task4.py. They were written during the deep-dive session and are now self-contained here (they read focus_180.json from this folder and no longer depend on any temp directory). Run with the repo on the path, e.g. PYTHONPATH=<repo> MPLBACKEND=Agg python scripts/<file>.

Script	What it does	Output folder
run_quarters.py	Token-quarter and feature-quarter ablations (zero/mean/resample), 180 samples	outputs/ijepa_task4_quarters/
run_deep.py	Q3 split, per-target-token, attention / anisotropy / prediction-lens (Run A, 180) and the full 1,000-sample quarter+Q3 sweep (Run B). Shared helpers other scripts import.	outputs/ijepa_task4_deep/expA_*, expB_*
run_resample_class.py	Resample donors drawn from same / nearest-other / farthest class	outputs/ijepa_task4_deep/resample_by_class_180/
run_task7_prototype.py	Quick Task-7 prototype: error-space PCA + precision-weighted Mahalanobis, per-token, full subset	outputs/ijepa_task7/
focus_180.json	The 9-class / 180-image focus subset used by the deep dives	—

Note: run_ijepa_task7.py (the fuller Task-7 implementation with a calibration split and both zero/centered Mahalanobis) is separate from the quick prototype above.