For reviewers โ scrub the cross-dataset evidence
Pick a dataset and a window โ watch HCIE vs every baseline
Every bar is the canonical Kalman-alone m_K readout (predict-before-observe, tie-aware), recomputed the same way on all six datasets. Click a dataset and a window. The honest story: HCIE is competitive, leads at the coldest window on some corpora, and trails the fully-trained deep models on dense data.
Junyi 2015 ยท first 5: HCIE m_K 0.689 โ trails (best 0.738) A fully-trained deep model leads here; HCIE stays competitive without training. The thesis contribution is the auditable loop + 8 structural axes, not topping every cell.
Readout flip (why the deck differs) The deployed 2-learner readout here is 0.560 vs canonical m_K 0.689. Slides built before the m_K switch may cite the deployed value โ this is the same CSEDM 0,673โ0,605 flip. The canonical bars above are the ones to cite.
Reproducible Canonical lagged-Kalman m_K, tie-aware, predict-before-observe โ identical method on all six datasets (validated CSEDM-exact vs cross_dataset_mK_unified.json). Source: coldstart_mK_all_datasets.json.