cs_check() now accounts for prediction error shared by all items of a family (the unseen family’s unknown effect, or the estimation error of a seen family’s effect). predict() returns this as sd_shared, estimated from the cross-validation and leave-one-family-out residuals, and cs_calibrate() carries it as prior_sd_shared. Previously a family unseen in training was falsely flagged 3.5-6% of the time at alpha = 0.01; it is now flagged at about the nominal rate. Seen families, detection of drifted families and calibration accuracy are unchanged. Calibrations without the new column behave exactly as before.
coldstart 0.1.0
Initial release.
Difficulty prediction from item features with honest seen/unseen-family uncertainty (cs_predictor()).
Grid-Bayes calibration with robust Student-t predicted priors (cs_calibrate()).