0.5.0

• Sparse lncRNA panel by L1 selection and L2 refit (scripts/16_sparse_panel.py): a few hundred genes are needed, not a few dozen; the 376-gene panel is listed.
• Bootstrap stability of the marker lncRNAs (scripts/14_marker_stability.py) and a direct measurement of the hospital-site signal within cancer types (scripts/15_site_signal.py).
• Calibration under shift and results by biopsy site on MET500 (lncpan.calibration, scripts/13_met500_calibration.py): temperature scaling fitted on TCGA barely helps on MET500; liver and lung biopsies are often called the host organ's cancer.
• Randomised APS and regularised adaptive prediction sets (RAPS) in lncpan.conformal, with the RAPS constants tuned on a split of the calibration set; lncpan predict --score raps. The deterministic APS score gave 14-type sets and empty sets to the most confident tumours; the randomised APS score covers 94% of the external metastases with one global threshold.