Leakage-Safe Evaluation for Multimodal Boiling Machine Learning: A Domain-Shift Baseline on BoilingBench

Random-window and pooled random splits can overstate the deployability of boiling machine-learning models when adjacent windows share a physical run or when preprocessing crosses the split boundary. We present a conservative, leakage-aware baseline evaluation on the public BoilingBench-Multimodal release. Per maintainer guidance (2026-09-18), the current public multimodal baseline identity is BoilingBench-3 (Cu-foam) and BoilingBench-4 (flat-Cu) from the NED3-002 lineage; NED3-007 is not part of this public release and is not used in task or split names here. The study freezes an archive revision (hanhuark/BoilingBench-Multimodal @ e4d977db, v0.1.0), documents a dataset-level leave-one-out protocol across BB-1--BB-4 as diagnostics, compares learned models with constant predictors via a pinned external CHF-Watch adapter (no vendoring), and preserves negative transfer as a result. Temporal heat-flux regression shows baseline-level BB-1 transfer and catastrophic BB-2 instability, while BB-3 and BB-4 are lower-error diagnostics. Visual bubble-area regression improves on base-test and new-facility evaluations but fails on flow external data. All performance numbers are provisional until independently reproduced by maintainers on repo-defined grouped splits with modality availability and label provenance documented. These are benchmark and failure-analysis results, not CHF safety validation: one processed run per BB-1--BB-4, heat flux is derived, chf_proxy is not confirmed CHF ground truth, public BB-3/4 lack continuous AE waveform exports, the private per-run CHF spreadsheet is excluded (published aggregates only), and any microchannel→BB-4 association remains inferred / awaiting maintainer confirmation.